Deep Learning Reconstruction to Improve the Quality of MR Imaging: Evaluating the Best Sequence for T-category Assessment in Non-small Cell Lung Cancer Patients

Abstract

Purpose

Deep learning reconstruction (DLR) has been recommended as useful for improving image quality. Moreover, compressed sensing (CS) or DLR has been proposed as useful for improving temporal resolution and image quality on MR sequences in different body fields. However, there have been no reports regarding the utility of DLR for image quality and T-factor assessment improvements on T2-weighted imaging (T2WI), short inversion time (TI) inversion recovery (STIR) imaging, and unenhanced- and contrast-enhanced (CE) 3D fast spoiled gradient echo (GRE) imaging with and without CS in comparison with thin-section multidetector-row CT (MDCT) for non-small cell lung cancer (NSCLC) patients. The purpose of this study was to determine the utility of DLR for improving image quality and the appropriate sequence for T-category assessment for NSCLC patients.

Methods

As subjects for this study, 213 pathologically diagnosed NSCLC patients who underwent thin-section MDCT and MR imaging as well as T-factor diagnosis were retrospectively enrolled. SNR of each tumor was calculated and compared by paired t-test for each sequence with and without DLR. T-factor for each patient was assessed with thin-section MDCT and all MR sequences, and the accuracy for T-factor diagnosis was compared among all sequences and thin-section CT by means of McNemar’s test.

Results

SNRs of T2WI, STIR imaging, unenhanced thin-section Quick 3D imaging, and CE-thin-section Quick 3D imaging with DLR were significantly higher than SNRs of those without DLR (P < 0.05). Diagnostic accuracy of STIR imaging and CE-thick- or thin-section Quick 3D imaging was significantly higher than that of thin-section CT, T2WI, and unenhanced thick- or thin-section Quick 3D imaging (P < 0.05).

Conclusion

DLR is thus considered useful for image quality improvement on MR imaging. STIR imaging and CE-Quick 3D imaging with or without CS were validated as appropriate MR sequences for T-factor evaluation in NSCLC patients.

Introduction

Lung cancer affects an estimated 2 million new patients each year and is associated with 1.76 million deaths per year, making it the leading cause of cancer-related death in the world.¹ Some investigators have suggested that multimodality treatment, including preoperative systemic therapy for preoperatively diagnosed locally advanced non-small cell lung cancer (NSCLC), may be useful because of the unfavorable prognosis for surgery alone.^2–4 This means that preoperative assessment of locoregional staging has become very important for management of NSCLC.

Although it was suggested in 1991 that unenhanced or contrast-enhanced (CE) MR imaging and angiography could be useful for mediastinal, chest wall or pulmonary vasculature invasions, unenhanced or CE thin-section multidetector-row CT (MDCT) with multiplanar reconstruction (MPR) images have been used since 2005 for differentiating T3 or T4 tumors from T1 or T2 tumors in routine clinical practice because of its potential which is superior to that of routine MDCT.^5–9 In view of these findings, there has been a continual need during the last decade for spatial, temporal, or contrast resolutions of MR imaging, while 3D fast spoiled gradient echo (GRE) sequences with or without fat suppression techniques have been suggested by all MR vendors as useful for various clinical purposes. Moreover, compressed sensing (CS) or deep learning reconstruction (DLR) have been proposed as useful for improving temporal resolution and image quality on MR sequences in different body fields.^10–13 However, there have been no reports regarding the utility of DLR for image quality and T-category assessment improvements on T2-weighted imaging (T2WI), short inversion time (TI) inversion recovery (STIR) imaging, and unenhanced- and CE-3D fast spoiled GRE imaging with and without CS in comparison with thin-section MDCT for NSCLC patients. Moreover, there have been no reports addressing question of the appropriate sequence for T-category evaluation of NSCLC patients using currently applied MR sequences. The purpose of this study was thus to determine the utility of DLR compared with that of thin-section MDCT for improving image quality on MR imaging and the appropriate sequence for T-category assessment of NSCLC patients.

Materials and Methods

This retrospective study was approved by the Institutional Review Board of Fujita Health University Hospital, is compliant with the Health Insurance Portability and Accountability Act, and written informed consent was waived. This study was technically and financially supported by Canon Medical Systems Corporation. Four of the authors are employees of Canon Medical Systems Corporation (K.Y., M.S., M.I., and M.Y.) but did not have control over any of the data used in this study.

Subjects

Between January 2020 and December 2022, 245 pathologically diagnosed NSCLC patients were retrospectively enrolled in this study. They underwent unenhanced or CE-thin-section MDCT and MR imaging on a 3T MR system at our hospitals less than 2 weeks {7±3 days (mean ± SD[standard deviation])} prior to the treatment. The exclusion criteria were: 1) no pathological T-category diagnosis of surgically treated NSCLC, 2) thin-section MDCT and MR imaging findings, which were not obtained for clinical T-category diagnosis by the tumor board of our hospital, 3) pregnancy or breast feeding, and 4) contraindication for MR imaging (pacemakers, ferromagnetic implants, etc.). Criterion 1) led to the exclusion of 7 patients, criterion 2) to that of 5 patients, criterion 3) to that of 2 pregnant and 5 breast feeding patients, and criterion 4) to that of 13 patients, 6 of who had a pacemaker and 7 who had ferromagnetic implants. None of the remaining 213 patients had been enrolled in any previously presented or published studies (Fig. 1). Details of patients’ characteristics are shown in Table 1.

Fig. 1.

Patient selection flow chart. A total of 213 subjects from 245 pathologically diagnosed NSCLC patients were selected as final study group in this study. MDCT, multidetector-row CT; NSCLC, non-small cell lung cancer; SD, standard deviation.

Table 1.

Patients’ characteristics

Participants and nodule characteristics		Units	Value
Gender	Men	Subjects	137/213 (64.4)
	Women		76/213 (35.6)
Age	Men	Years (mean ± SD [range])	71 ± 8 (47–83)
	Women		69 ± 10 (42–85)
Mean smoking history		Pack-years (mean ± SD [range])	34 ± 10 (20–154)
Histological subtypes	MIA	Cases	12/213 (5.6)
		Long axis diameter (mm) (mean ± SD [range])	13.2 ± 4.1 (5–19)
	Invasive adenocarcinoma	Cases	103/213 (48.4)
		Long axis diameter (mm) (mean ± SD [range])	21.3 ± 7.1 (5–54)
	Adenocarcinoma with other subtypes	Cases	68/213 (31.9)
		Long axis diameter (mm) (mean ± SD [range])	26.8 ± 9.2 (12–84)
	Squamous cell carcinoma	Cases	18/213 (8.5)
		Long axis diameter (mm) (mean ± SD [range])	29.4 ± 8.9 (15–79)
	Large cell carcinoma	Cases	8/213 (3.7)
		Long axis diameter (mm) (mean ± SD [range])	27.3 ± 8.1 (13–55)
	Adenosquamous cell carcinoma	Cases	4/213 (1.9)
		Long axis diameter (mm) (mean ± SD [range])	18.5 ± 7.8 (8–39)
T-category	T1a	Cases	13
	T1b	Cases	55
	T1c	Cases	38
	T2a	Cases	25
	T2b	Cases	21
	T3	Cases	34
	T4	Cases	27

MIA, Minimally invasive adenocarcinoma; SD, standard deviation.

Radiological examinations

Chest MR imaging of each patient was performed using a 3 tesla (T) MRI system (Vantage Centurian; Canon Medical Systems, Otawara, Tochigi, Japan) with a 16-element phased-array surface coil (Atlas SPEEDER Body and Atlas SPEEDER Spine, Canon Medical Systems) combined with parallel imaging capability (SPEEDER, Canon Medical Systems). The SPEEDER reduction factor was the acceleration factor used for parallel imaging. Chest MRI for each patient was obtained with the following six sequences: 1) electrocardiogram (ECG)-triggered T2-weighted fast spin-echo (FSE) sequence, 2) STIR fast advanced spin-echo (FASE), which was a sequentially reordered half-Fourier reconstructed FSE with a short TE and long echo train length, 3) unenhanced and thick-section fast and segmented 3D T1-weighted spoiled GRE sequence (Quick 3D, Canon Medical) using a double fat suppression (DFS) RF pulse technique (DFS), 4) unenhanced and thin-section Quick 3D with CS, 5) thick-section Quick 3D with contrast media, and 6) thin-section Quick 3D with CS and contrast media. For each CE MR imaging, a standard dose (0.05 mmoL/kg BW) of contrast medium (Magnescope; Guerbet, Tokyo, Japan) was administered intravenously, and each contrast enhanced study was started 1 min after contrast medium administration. All CE MR imaging were obtained in random order. All MR data were then reconstructed with and without DLR (Advanced intelligent Clear-IQ Engine [AiCE]; Canon Medical Systems). Details of the six sequences used in this study are listed in Table 2. Moreover, details of CS and DLR had been stated in the past literatures.^10–13

Table 2.

Details of sequences

Sequence	ECG-triggered T2WI	ECG-triggered STIR FASE	Unenhanced thick-section Quick 3D with DFS technique	Unenhanced thin-section Quick 3D with DFS technique and CS	Contrast-enhanced thick-section Quick 3D with DFS technique	Contrast-enhanced thin-section Quick 3D with DFS technique and CS
FOV	320 × 400
Coil	Atlas Spine (16 elements) and Atlas Body (16 elements)
TR (ms)	2–3 <R-R>	2–3 <R-R>	3.3	3.3	3.3	3.3
TE (ms)	80	10	1.1	1.1	1.1	1.1
Black Blood TI (ms)	600	600	N/A
STIR TI (ms)	N/A	200	N/A
Flip angle (degree)	90/160	90/160	9
ETL	44	24	N/A
Parallel imaging method	SPEEDER	SPEEDER	SPEEDER	Compressed SPEEDER	SPEEDER	Compressed SPEEDER
Reduction factor
Phase direction	1.9	1.9	1.5	4	1.5	4
Slice direction	N/A	N/A	N/A	2.2	N/A	2.2
NEX	1
Section thickness (mm)	5	5	5	1	5	1
Slice gap (mm)	1.5	1.5	0	0	0	0
Slice number	40–44	40–44	50–55	250–275	50–55	250–275
Matrix	192 × 256	192 × 256	144 × 320	320 × 400	144 × 320	320 × 400
Reconstruction matrix	384 × 512	384 × 512	480 × 640	640 × 800	480 × 640	640 × 800
Breath hold (sec)	15–16s × 10–12 times	15–16s × 10–12 times	20–22	19–23	20–22	19–23
Reconstruction method	with and without DLR

CS, compressed sensing; DFS, double fat suppression; DLR, deep learning reconstruction; ECG, electrocardiogram; ETL, echo train length; NEX, number of excitations; STIR FASE, short inversion time inversion-recovery fast advanced spin-echo; TI, Inversion time; T2WI, T2-weighted imaging.

Unenhanced and CE-MDCT examinations were performed using three 320-detector row CT scanners (Aquilion ONE; Canon Medical Systems) as a 64-detector row helical CT mode. The examinations included three separate acquisitions: chest acquisition from lung apex to diaphragm at the end of suspended inspiration, abdomen acquisition from the diaphragm to the anus at the end of suspended inspiration, and neck acquisition from skull base to lung apex. For patients without contraindication to contrast media administration, the images were obtained after intravenous administration of 100 ml of contrast material (Iopamiron 300; Bayer Pharma, Osaka, Japan) using an automatic infusion system (Dual Shoot GX7; Nemoto Kyorindo, Tokyo, Japan) at a rate of 3 mL/s. The mean CTDI_vol was 34.8mGy (range, 15.1–51.0mGy). The estimated dose per length, calculated as the product of CTDI_vol and scan length, ranged from 496.5 to 3026.5mGy cm. The estimated effective dose ranged from 6.96 to 42.38 mSv. Next, chest CT images were reconstructed as contiguous axial images with a section thickness of 5 mm, contiguous axial CT images with a section thickness of 1 mm, and as contiguous sagittal and coronal MPR images, all of which were obtained with a hybrid-type reconstruction algorithm (adaptive iterative dose reduction 3D enhance [AIDR 3D enhance]; Canon Medical Systems) as well as with lung and standard kernels (FC52 and FC13; Canon Medical Systems).

Standard references for T-category

For all surgically treated NSCLC patients, all pathologic specimens stained with the hematoxylin-eosin stain and/or the elastic van Gieson stain were reviewed by board certified pathologists with more than 10 years of experience who were not included in this study and unaware of the findings of any of the radiological examinations as specified by the Union Internationale Contre le Cancer (UICC), 8th edition.^14,15 In contrast to the procedure for unoperated NSCLC patients, T-categories were assessed by tumor boards, which consisted of board-certified surgeons, pulmonologists, radiologists, and pathologists with more than 10 years of experience who were not included in this study.

Image analysis

Quantitative and Qualitative Image Quality Evaluation

All quantitative and qualitative assessments of image quality and qualitative T-category evaluation of each method were performed with the aid of a commercially available picture archiving and communication system (PACS) (RapideyeCore; Canon Medical Systems).

For quantitative image quality assessment, ROIs with the same diameter were placed over each tumor, the intercostal or trapezius muscles on both sides of the same slice plane and the trachea by a board-certified chest radiologist (Y. Ohno.) with 27 years of experience. For quantitative image quality assessment, SNRs of tumor and muscle and contrast-to-noise ratio (CNR) between the tumor and muscle on each MR protocol were calculated by using the following formulas derived from previous reports^10–13

$$\mathrm{S}\mathrm{N}{\mathrm{R}}_{\mathrm{t}\mathrm{u}\mathrm{m}\mathrm{o}\mathrm{r}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{u}\mathrm{s}\mathrm{c}\mathrm{l}\mathrm{e}}=\mathrm{S}{\mathrm{I}}_{\mathrm{t}\mathrm{u}\mathrm{m}\mathrm{o}\mathrm{r}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{u}\mathrm{s}\mathrm{c}\mathrm{l}\mathrm{e}}/\mathrm{S}{\mathrm{D}}_{\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{h}\mathrm{e}\mathrm{a}}$$ [1]

where SI_tumor or SI_muscle is the averaged signal intensity of the tumor or muscle within the ROI, and SD_trachea is the average standard deviation of the trachea within the ROI.

And for CNR:

$$\mathrm{C}\mathrm{N}\mathrm{R}=\mathrm{S}\mathrm{N}{\mathrm{R}}_{\mathrm{t}\mathrm{u}\mathrm{m}\mathrm{o}\mathrm{r}}-\mathrm{S}\mathrm{N}{\mathrm{R}}_{\mathrm{m}\mathrm{u}\mathrm{s}\mathrm{c}\mathrm{l}\mathrm{e}}$$ [2]

For qualitative assessment of image quality, the same chest radiologist (Y. Ohno) and another board-certified chest with 30 years of experience (D.T.) independently and visually evaluated overall image quality, artifact incidence, and diagnostic confidence level of each MR protocol by using a 5-point scoring system. For this study, overall image quality was rated as: 1, poor; 2, fair; 3, moderate; 4, good; and 5, excellent. Artifacts were scored as: 1, no visualization of artifacts; 2, visualization of a few artifacts; 3, visualization of some artifacts; 4, visualization of several artifacts; and 5, visualization of significant numbers of artifacts. As artifacts, motion, ringing, pulsation, sensitivity encoding (SENSE)-specific, and susceptibility artifacts were evaluated in this study. All final visual scores for every patient were determined by consensus of the two readers.

Qualitative T-Category Evaluation

The same board-certified chest radiologists who evaluated overall image quality also evaluated the probability of the occurrence of T3 or T4 cases by using the following 5-point visual scoring system: 1, indefinite; 2, probably indefinite, 3, uncertain; 4, probably definite; and 5, definite. The diagnostic criteria for mediastinal or chest wall invasions were based on those published in the literature.^10,16–22 For mediastinal and chest wall invasions on MDCT and for each MR protocol, the criteria were: (a) contact with pleura over more than 3 cm, (b) obtuse angle, (c) extending to fat plane, (d) obliteration of fat plane, or (e) abnormal signal intensity or enhancement due to tumor invasion within mediastinal structure and chest wall.^20–26 The T-category obtained with each method was based on the definition agreed upon by of the Union Internationale Contre le Cancer (UICC), 8th edition.^14,15

Statistical analysis

Quantitative and Qualitative Image Quality Comparisons

To determine the capability of DLR for each sequence, SNR and CNR obtained with each sequence with and without DLR were compared by using the paired t-test.

For a comparison of qualitative image quality, interobserver agreements for overall image quality, artifacts, and diagnostic confidence level for each method were assessed by means of weighted kappa statistics. Wilcoxon’s signed rank test was then used for a comparison of overall image quality and artifact level attained by each sequence with and without DLR.

Qualitative T-category Evaluation Capability Comparison

Receiver operating characteristic (ROC) analyses were performed to compare the diagnostic capability of all the methods for differentiation of T3 or T4 from T1 or T2. Next, sensitivity, specificity, and accuracy for distinguishing T3 or T4 from T1 or T2 attained by all methods were compared by means of Cochran’s Q test.

Kappa statistics were used to determine interobserver agreement for T-category evaluation for each method as compared with standard reference. Diagnostic accuracy for T-category was then compared among all sequences and thin-section CT with the aid of Cochran’s Q test.

To assess diagnostic performance of each method for distinguishing T4- from T3-categories, kappa statistics were used to determine interobserver agreement for distinguishing T4- from T3-categories for each method as compared with standard reference, and diagnostic performance was then compared among all methods with the aid of Cochran’s Q test.

All agreements were considered slight for κ < 0.21, fair for κ = 0.21–0.40, moderate for κ = 0.41–0.60, substantial for κ = 0.61–0.80, and almost perfect for κ = 0.81–1.00.²³ A P value less than 0.05 was considered significant for all statistical analyses. All statistical analyses were performed using JMP version 14 (SAS Institute Japan, Toyko, Japan).

Results

The group consisting of 213 qualifying NSCLC patients (mean age 70 years ± 9 [SD], 137 men) constituted the eventual study group for evaluations. In this group, 132 patients underwent surgical treatment and 81 were treated non-surgically. Treatment of the former consisted of 101 lobectomies, 13 partial lobectomies, 9 segmentectomies, 4 bilobectomies, 3 sleeve lobectomies, and 2 pneumonectomies. All NSCLC patients were pathologically diagnosed as having invasive lepidic adenocarcinomas (n = 103), adenocarcinomas with other subtypes (n = 68), squamous cell carcinomas (n = 18), minimally invasive adenocarcinomas (MIAs, n = 12), large cell carcinomas (n = 8), and adenosquamous cell carcinomas (n = 4). In terms of T-categories, 13 patients were classified T1a cases, 55 as T1b cases, 38 as T1c cases, 25 as T2a cases, 21 as T2b cases, 34 as T3 cases, and 27 as T4 cases. Of the T3 cases, 33 showed chest wall, phrenic nerve or pericardial invasions, and of the T4 cases, 24 had mediastinal, diaphragm, cardiac, great vessel, tracheal, esophageal, or vertebral invasions.

SNRs of tumors and muscles and CNRs between tumors and muscles for each sequence are shown in Table 3. SNRs of tumor and muscle of T2WI, STIR imaging, unenhanced thin-section Quick 3D imaging, and CE-thin-section Quick 3D imaging with DLR were significantly higher than those of sequences without DLR (P < 0.05). Representative case for image analysis is shown in Fig. 2.

Table 3.

SNRs of tumor and muscle and CNR between tumor and muscle determined with each method

Method	SNR of tumor	SNR of muscle	CNR between tumor and muscle
ECG-gated T2WI reconstructed without DLR	22.2 ± 12.1*	23.0 ± 17.4*	3.5 ± 8.4
ECG-gated T2WI reconstructed with DLR	34.5 ± 23.2	33.5 ± 22.9	4.1 ± 10.1
ECG-gated STIR reconstructed without DLR	25.2 ± 11.3*	11.1 ± 3.8*	14.9 ± 9.0
ECG-gated STIR reconstructed with DLR	42.2 ± 21.7	18.5 ± 7.3	18.2 ± 12.2
Unenhanced thick-section Quick 3D without DLR	18.7 ± 11.1	26.0 ± 14.4	−3.7 ± 5.1
Unenhanced thick-section Quick 3D with DLR	22.0 ± 13.4	31.1 ± 18.5	−4.2 ± 5.7
Unenhanced thin-section Quick 3D without DLR	7.5 ± 3.6*	10.3 ± 4.3*	−2.4 ± 3.1
Unenhanced thin-section Quick 3D with DLR	11.8 ± 5.8	17.3 ± 8.9	−4.3 ± 5.9
CE-thick-section Quick 3D without DLR	21.3 ± 16.3	18.9 ± 11.1	−0.4 ± 6.8
CE-thick-section Quick 3D with DLR	23.2 ± 17.9	20.4 ± 12.0	−0.5 ± 7.5
CE-thin-section Quick 3D without DLR	9.9 ± 4.9*	9.5 ± 4.1*	−0.4 ± 4.4
CE-thin-section Quick 3D with DLR	15.6 ± 6.3	15.7 ± 7.9	−1.1 ± 7.0

CE, contrast-enhanced; CNR, contrast-to-noise ratio, DLR, deep learning reconstruction; ECG, electrocardiogram, STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

*: Significantly lower than each sequence with DLR (P < 0.05).

Fig. 2.

A 71-year-old male patient with adenocarcinoma, which was diagnosed as T1c, in the left upper lobe and mediastinal lymph node metastasis (N2 disease). The application of DLR resulted in improvement of image noise for T2WI, STIR image, unenhanced-, and CE-thin-section Quick 3D images. However, image noise improvement for unenhanced- and CE-thick-section Quick 3D images was somewhat less than that for other images. Each imaging method was scored as 1 and diagnosed as T1c. This case was evaluated as true-negative case in this study. CE, contrast-enhanced; DLR, deep learning reconstruction; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

Interobserver agreements between two investigators and comparisons of overall image and artifacts between each sequence with and without DLR are shown in Table 4 and 5. All interobserver agreements for overall image quality (0.64 ≤ κ ≤ 0.82, P < 0.0001) and artifacts (0.62 ≤ κ ≤ 0.72, P < 0.0001) were determined as significantly substantial or almost perfect. The application of DLR median resulted in significantly improved overall image quality and artifacts of T2WI, STIR imaging, unenhanced thin-section Quick 3D imaging, CE-thick-section Quick 3D imaging, and CE-thin-section Quick 3D imaging (P < 0.05).

Table 4.

Interobserver agreements between two investigators and comparison of overall image quality attained with each sequence with and without DLR

Method	Investigators	Overall image quality					Interobserver agreements		Median
		1	2	3	4	5	Kappa value	P value	(IQR)
ECG-gated T2WI reconstructed without DLR	Investigator 1	0	0	11	89	113	0.77	< 0.0001	5*
	Investigator 2	0	0	27	85	101			(4–5)
ECG-gated T2WI reconstructed with DLR	Investigator 1	0	0	0	43	170	0.66	< 0.0001	5
	Investigator 2	0	0	5	61	147			(5–5)
ECG-gated STIR reconstructed without DLR	Investigator 1	0	0	117	90	6	0.64	< 0.0001	3*
	Investigator 2	0	11	129	73	0			(3–4)
ECG-gated STIR reconstructed with DLR	Investigator 1	0	0	0	101	112	0.65	< 0.0001	5
	Investigator 2	0	0	16	109	88			(4–5)
Unenhanced thick-section Quick 3D without DLR	Investigator 1	117	49	16	20	11	0.81	< 0.0001	4
	Investigator 2	113	56	12	19	13			(4–5)
Unenhanced thick-section Quick 3D with DLR	Investigator 1	117	49	15	21	11	0.82	< 0.0001	4
	Investigator 2	113	56	10	21	13			(4–5)
Unenhanced thin-section Quick 3D without DLR	Investigator 1	117	49	12	24	11	0.79	< 0.0001	3*
	Investigator 2	113	58	9	20	13			(2–3)
Unenhanced thin-section Quick 3D with DLR	Investigator 1	117	49	9	27	11	0.81	< 0.0001	3
	Investigator 2	113	56	9	22	13			(3–4)
CE-thick-section Quick 3D without DLR	Investigator 1	117	46	12	25	13	0.79	< 0.0001	4*
	Investigator 2	112	55	15	21	10			(4–5)
CE-thick-section Quick 3D with DLR	Investigator 1	117	46	10	23	17	0.82	< 0.0001	5
	Investigator 2	110	57	12	17	17			(4–5)
CE-thin-section Quick 3D without DLR	Investigator 1	117	42	6	24	24	0.80	< 0.0001	4*
	Investigator 2	114	45	14	20	20			(3–4)
CE-thin-section Quick 3D with DLR	Investigator 1	117	42	3	18	33	0.82	< 0.0001	5
	Investigator 2	115	44	7	28	19			(4–5)

CE, contrast-enhanced; DLR, deep learning reconstruction; ECG, electrocardiogram; IQR, interquartile range; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

*: Significantly lower than each sequence with DLR (P < 0.05).

Table 5.

Interobserver agreements between two investigators and comparison of artifacts for sequences with and without DLR

Method	Investigators	Artifact					Interobserver agreements		Median
		1	2	3	4	5	Kappa value	P value	(IQR)
ECG-gated T2WI reconstructed without DLR	Investigator 1	113	89	11	0	0	0.66	< 0.0001	1*
	Investigator 2	83	107	11	0	0			(1–2)
ECG-gated T2WI reconstructed with DLR	Investigator 1	170	43	0	0	0	0.66	< 0.0001	1
	Investigator 2	146	67	0	0	0			(1–1)
ECG-gated STIR reconstructed without DLR	Investigator 1	6	90	117	0	0	0.62	< 0.0001	3*
	Investigator 2	0	72	129	12	0			(2–3)
ECG-gated STIR reconstructed with DLR	Investigator 1	112	101	0	0	0	0.62	< 0.0001	1
	Investigator 2	76	131	6	0	0			(1–2)
Unenhanced thick-section Quick 3D without DLR	Investigator 1	59	111	43	0	0	0.67	< 0.0001	2
	Investigator 2	35	129	37	12	0			(1–2)
Unenhanced thick-section Quick 3D with DLR	Investigator 1	89	104	20	0	0	0.67	< 0.0001	2
	Investigator 2	65	110	38	0	0			(1–2)
Unenhanced thin-section Quick 3D without DLR	Investigator 1	0	33	126	48	6	0.65	< 0.0001	3*
	Investigator 2	0	15	138	36	24			(3–4)
Unenhanced thin-section Quick 3D with DLR	Investigator 1	27	77	77	32	0	0.72	< 0.0001	3
	Investigator 2	15	77	77	38	6			(2–3)
CE-thick-section Quick 3D without DLR	Investigator 1	95	96	22	0	0	0.68	< 0.0001	2*
	Investigator 2	71	108	28	6	0			(1–2)
CE-thick-section Quick 3D with DLR	Investigator 1	122	91	0	0	0	0.63	< 0.0001	1
	Investigator 2	92	109	12	0	0			(1–1)
CE-thin-section Quick 3D without DLR	Investigator 1	42	106	59	6	0	0.70	< 0.0001	2*
	Investigator 2	30	100	65	12	6			(2–3)
CE-thin-section Quick 3D with DLR	Investigator 1	111	79	17	6	0	0.68	< 0.0001	1
	Investigator 2	81	97	29	6	0			(1–2)

CE, contrast-enhanced; DLR, deep learning reconstruction; ECG, electrocardiogram; IQR, interquartile range; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging. *: Significantly lower than each sequence with DLR (P < 0.05).

Interobserver agreement for the probability of occurrence of T3 or T4 cases determined with each method is shown in Table 6. Interobserver agreements for all methods were rated substantial or almost perfect (0.72 ≤ κ ≤ 0.82, P < 0.0001).

Table 6.

Interobserver agreements for the probability of T3 or T4 cases occurring for each method

Method	Investigators	Probability for T3 or T4					Interobserver agreements
		1	2	3	4	5	Kappa value	P value
Thin-section MDCT	Investigator 1	115	48	16	20	14	0.72	< 0.0001
	Investigator 2	112	52	16	21	12
ECG-gated T2WI reconstructed without DLR	Investigator 1	117	45	12	24	15	0.76	< 0.0001
	Investigator 2	116	48	14	21	14
ECG-gated T2WI reconstructed with DLR	Investigator 1	117	45	9	24	18	0.79	< 0.0001
	Investigator 2	115	49	13	24	12
ECG-gated STIR reconstructed without DLR	Investigator 1	117	42	15	17	22	0.80	< 0.0001
	Investigator 2	111	51	13	23	15
ECG-gated STIR reconstructed with DLR	Investigator 1	117	45	8	20	23	0.81	< 0.0001
	Investigator 2	112	48	13	23	17
Unenhanced thick-section Quick 3D without DLR	Investigator 1	117	49	16	20	11	0.81	< 0.0001
	Investigator 2	113	56	12	19	13
Unenhanced thick-section Quick 3D with DLR	Investigator 1	117	49	15	21	11	0.82	< 0.0001
	Investigator 2	113	56	10	21	13
Unenhanced thin-section Quick 3D without DLR	Investigator 1	117	49	12	24	11	0.79	< 0.0001
	Investigator 2	113	58	9	20	13
Unenhanced thin-section Quick 3D with DLR	Investigator 1	117	49	9	27	11	0.81	< 0.0001
	Investigator 2	113	56	9	22	13
CE-thick-section Quick 3D without DLR	Investigator 1	117	46	12	25	13	0.79	< 0.0001
	Investigator 2	112	55	15	21	10
CE-thick-section Quick 3D with DLR	Investigator 1	117	46	10	23	17	0.82	< 0.0001
	Investigator 2	110	57	12	17	17
CE-thin-section Quick 3D without DLR	Investigator 1	117	42	6	24	24	0.80	< 0.0001
	Investigator 2	114	45	14	20	20
CE-thin-section Quick 3D with DLR	Investigator 1	117	42	3	18	33	0.82	< 0.0001
	Investigator 2	115	44	7	28	19

CE, contrast-enhanced; DLR, deep learning reconstruction; ECG, electrocardiogram; MDCT, multi-detector row CT; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

Results of a comparison of ROC analysis and diagnostic performance for differentiating T3 or T4 tumor from T1 or T2 tumor by all methods are shown in Table 7. Areas under the curves (AUCs) were significantly larger, while sensitivities and accuracies of STIR imaging with and without DLR and CE-thin-section Quick 3D imaging with and without DLR were significantly higher than those of thin-section CT, T2WI with and without DLR, unenhanced thick-section, and thin-section Quick 3D imaging with and without DLR (P < 0.05).

Table 7.

Results of ROC analysis and diagnostic performance compared among all methods

Method	AUC	Threshold value	SE (%)	SP (%)	PPV (%)	NPV (%)	AC (%)
Thin-section MDCT	0.97*, **, ***, ****	3	87.7 (50/57)	100 (156/156)	100 (50/50)	95.7 (156/163)	96.7 (206/213)
ECG-gated T2WI reconstructed without DLR	0.97*, **, ***, ****	3	89.5 (51/57)	100 (156/156)	100 (51/51)	96.3 (156/162)	97.2 (207/213)
ECG-gated T2WI reconstructed with DLR	0.97*, **, ***, ****	3	89.5 (51/57)	100 (156/156)	100 (51/51)	96.3 (156/162)	97.2 (207/213)
ECG-gated STIR reconstructed without DLR	0.98	3	96.5 (55/57)	100 (156/156)	100 (55/55)	98.7 (156/158)	99.1 (211/213)
ECG-gated STIR reconstructed with DLR	0.98	3	96.5 (55/57)	100 (156/156)	100 (55/55)	98.7 (156/158)	99.1 (211/213)
Unenhanced thick-section Quick 3D without DLR	0.96*, **, ***, ****	3	82.4 (47/57) *, **, ***, ****	100 (156/156)	100 (47/47)	94.0 (156/166)	95.3 (203/213) *, **, ***, ****
Unenhanced thick-section Quick 3D with DLR	0.96*, **, ***, ****	3	82.4 (47/57) *, **, ***, ****	100 (156/156)	100 (47/47)	94.0 (156/166)	95.3 (203/213) *, **, ***, ****
Unenhanced thin-section Quick 3D without DLR	0.96*, **, ***, ****	3	82.4 (47/57) *, **, ***, ****	100 (156/156)	100 (47/47)	94.0 (156/166)	95.3 (203/213) *, **, ***, ****
Unenhanced thin-section Quick 3D with DLR	0.96*, **, ***, ****	3	82.4 (47/57) *, **, ***, ****	100 (156/156)	100 (47/47)	94.0 (156/166)	95.3 (203/213) *, **, ***, ****
CE-thick-section Quick 3D without DLR	0.97*, **, ***, ****	3	87.7 (50/57)	100 (156/156)	100 (50/50)	98.1 (156/159)	96.7 (206/213)
CE-thick-section Quick 3D with DLR	0.97*, *****, ****	3	87.7 (50/57)	100 (156/156)	100 (50/50)	98.1 (156/159)	96.7 (206/213)
CE-thin-section Quick 3D without DLR	0.98	3	94.7 (54/57)	100 (156/156)	100 (50/50)	98.1 (156/159)	98.6 (210/213)
CE-thin-section Quick 3D with DLR	0.98	3	94.7 (54/57)	100 (156/156)	100 (50/50)	98.1 (156/159)	98.6 (210/213)

AC, accuracy; AUC, area under the curve; CE, contrast-enhanced; DFS, double fat suppression; DLR, deep learning reconstruction; ECG, electrocardiogram; MDCT, multi-detector row CT; NPV, negative predictive value; PPV, positive predictive value; SE, sensitivity; SP, specificity; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

*: Significant difference with ECG-gated STIR reconstructed without DLR (P < 0.05).

**: Significant difference with ECG-gated STIR reconstructed with DLR (P < 0.05).

***: Significant difference with CE-thin-section Quick 3D using DFS without DLR (P < 0.05).

****: Significant difference with CE-thin-section Quick 3D using DFS with DLR (P < 0.05).

Interobserver agreements for and diagnostic accuracy of T-category evaluation for all the methods are shown in Table 8. Interobserver agreements for all methods were almost perfect (0.81 ≤ κ ≤ 0.93). Diagnostic accuracy of STIR imaging with and without DLR, CE-thick-section Quick 3D imaging with and without DLR, and CE-thin-section Quick 3D imaging with and without DLR was significantly higher than that of thin-section CT, T2WI with and without DLR, unenhanced thick-section Quick 3D imaging with and without DLR, and unenhanced thin-section Quick 3D imaging with and without DLR (P < 0.05).

Table 8.

Interobserver agreements and diagnostic accuracy of all methods for T-category evaluation

Method		T-category evaluation							Interobserver agreement		Accuracy (%)
	Undetected	T1a (n = 13)	T1b (n = 55)	T1c (n = 38)	T2a (n = 25)	T2b (n = 21)	T3 (n = 34)	T4 (n = 27)	Kappa value	P value
Thin-section MDCT	0	7	58	38	30	27	31	22	0.81	< 0.0001	84.0 (179/213) *, **, ***, ****, *****, ******
ECG-gated T2WI reconstructed without DLR	7	0	55	43	31	22	33	22	0.81	< 0.0001	84.0 (179/213) *, **, ***, ****, *****, ******
ECG-gated T2WI reconstructed with DLR	7	0	55	43	29	22	34	23	0.83	< 0.0001	85.4(182/213) *, **, ***, ****, *****, ******
ECG-gated STIR reconstructed without DLR	4	3	58	41	27	20	34	26	0.89	< 0.0001	90.6 (193/213)
ECG-gated STIR reconstructed with DLR	3	4	58	41	27	20	34	26	0.89	< 0.0001	91.1 (194/213)
Unenhanced thick-section Quick 3D without DLR	12	1	53	42	26	27	27	25	0.82	< 0.0001	84.5 (180/213) *, **, ***, ****, *****, ******
Unenhanced thick-section Quick 3D with DLR	12	1	53	42	26	27	27	25	0.82	< 0.0001	84.5 (180/213) *, **, ***, ****, *****, ******
Unenhanced thin-section Quick 3D without DLR	12	1	53	41	27	25	29	25	0.83	< 0.0001	85.9 (183/213) *, **, ***, ****, *****, ******
Unenhanced thin-section Quick 3D with DLR	11	2	53	41	27	25	29	25	0.84	< 0.0001	86.4 (184/213) *, **, ***, ****, *****, *******
CE-thick-section Quick 3D without DLR	3	10	53	42	26	24	30	25	0.89	< 0.0001	91.1(194/213)
CE-thick-section Quick 3D with DLR	3	10	53	42	26	24	30	25	0.89	< 0.0001	91.1(194/213)
CE-thin-section Quick 3D without DLR	2	11	53	41	26	21	32	27	0.90	< 0.0001	92.5 (197/213)
CE-thin-section Quick 3D with DLR	0	13	53	41	26	21	31	28	0.93	< 0.0001	94.8 (200/213)

AC, accuracy; CE, contrast-enhanced; DFS, double fat suppression; DLR, deep learning reconstruction; ECG, electrocardiogram; MDCT, multi-detector row CT; NPV, negative predictive value; PPV, positive predictive value; SE; sensitivity; SP, specificity; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

*: Significant difference with STIR imaging without DLR (P < 0.05).

**: Significant difference with STIR imaging with DLR (P < 0.05).

***: Significant difference with CE-thick-section Quick 3D using DFS without DLR (P < 0.05).

****: Significant difference with CE-thick-section Quick 3D using DFS with DLR (P < 0.05).

*****: Significant difference with CE-thin-section Quick 3D using DFS without DLR (P < 0.05).

******: Significant difference with CE-thin-section Quick 3D using DFS with DLR (P < 0.05).

Interobserver agreements for and diagnostic performance of all methods for distinguishing T4- from T3-catergories are shown in Table 9. Interobserver agreements for all methods were substantial (0.62 ≤ κ ≤ 0.79). Diagnostic accuracy of STIR imaging with and without DLR and CE-thin-section Quick 3D imaging with and without DLR was significantly higher than that of unenhanced thick-section Quick 3D imaging with and without DLR and unenhanced thin-section Quick 3D imaging with and without DLR (P < 0.05).

Table 9.

Interobserver agreements and diagnostic performance of all methods for distinguishing T4- from T3-catergory

Method	T-category evaluation			Interobserver agreement		Sensitivity (%)	Specificity (%)	Accuracy (%)
	Undetected or less than T3 (n = 0)	T3 (n = 34)	T4 (n = 27)	Kappa value	P value
Thin-section MDCT	11	31	19	0.69	< 0.0001	70.4 (19/27)	91.2 (31/34)	82.0 (50/61)
ECG-gated T2WI reconstructed without DLR	10	29	22	0.72	< 0.0001	81.5 (22/27)	85.3 (29/34)	83.6 (51/61)
ECG-gated T2WI reconstructed with DLR	10	29	22	0.72	< 0.0001	81.5 (22/27)	85.3 (29/34)	83.6 (51/61)
ECG-gated STIR reconstructed without DLR	7	31	23	0.79	< 0.0001	85.2 (23/27)	91.2 (31/34)	88.5 (54/61)
ECG-gated STIR reconstructed with DLR	7	31	23	0.79	< 0.0001	85.2 (23/27)	91.2 (31/34)	88.5 (54/61)
Unenhanced thick-section Quick 3D without DLR	11	27	20	0.62	< 0.0001	74.1 (20/27)	79.4 (27/34)	77.0 (47/61) *, **, *****, ******
Unenhanced thick-section Quick 3D with DLR	14	27	20	0.62	< 0.0001	74.1 (20/27)	79.4 (27/34)	77.0 (47/61) *, **, *****, ******
Unenhanced thin-section Quick 3D without DLR	14	27	20	0.62	< 0.0001	74.1 (20/27)	79.4 (27/34)	77.0 (47/61) *, **, *****, ******
Unenhanced thin-section Quick 3D with DLR	14	27	20	0.62	< 0.0001	74.1 (20/27)	79.4 (27/34)	77.0 (47/61) *, **, *****, ******
CE-thick-section Quick 3D without DLR	11	30	20	0.69	< 0.0001	74.1 (20/27)	88.2 (30/34)	82.0 (50/61)
CE-thick-section Quick 3D with DLR	11	30	20	0.69	< 0.0001	74.1 (20/27)	88.2 (30/34)	82.0 (50/61)
CE-thin-section Quick 3D without DLR	7	30	24	0.79	< 0.0001	88.9 (24/27)	88.2 (30/34)	88.5 (54/61)
CE-thin-section Quick 3D with DLR	7	30	24	0.79	< 0.0001	88.9 (24/27)	88.2 (30/34)	88.5 (54/61)

AC, accuracy; CE, contrast-enhanced; DFS, double fat suppression; DLR, deep learning reconstruction; ECG, electrocardiogram; MDCT, multi-detector row CT; NPV, negative predictive value; PPV, positive predictive value; SE, sensitivity; SP, specificity; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

*: Significant difference with STIR imaging without DLR (P < 0.05).

**: Significant difference with STIR imaging with DLR (P < 0.05).

***: Significant difference with CE-thick-section Quick 3D using DFS without DLR (P < 0.05).

****: Significant difference with CE-thick-section Quick 3D using DFS with DLR (P < 0.05).

*****: Significant difference with CE-thin-section Quick 3D using DFS without DLR (P < 0.05).

******: Significant difference with CE-thin-section Quick 3D using DFS with DLR (P < 0.05).

Representative case for T-factor evaluation is shown in Fig. 3.

Fig. 3.

A 75-year-old female patient with adenocarcinoma in the left upper lobe featuring aortic invasion and classified as a T4 case based on pathological examination result. On thin-section CT, mediastinal fat between tumor and aorta was observed, and mediastinal invasion was scored as 3. This case was assessed as true-positive for mediastinal invasion, even though this case was evaluated as T3. On T2WI with and without DLR, mediastinal fat between tumor and aorta was observed, and mediastinal invasion was scored as 2. This case was assessed as false-negative for mediastinal invasion. STIR images with and without DLR showed high signal intensity at the aortic wall and the presence of mediastinal, and aortic invasions was suspected, leading to a score of 5. This case was true-positive and evaluated as T4. Unenhanced and CE-thick- and thin-section Quick 3D images with and without DLR showed no mediastinal fat signal between tumor and aorta. As for assessment of mediastinal invasion, it was scored as 3 on unenhanced thick-section Quick 3D with and without DLR, as 4 on unenhanced thin-section Quick 3D with and without DLR and CE-thin-section Quick 3D without DLR, and as 5 on CE-thin-section Quick 3D with DLR. All images were evaluated as true-positive and T4. CE, contrast-enhanced; DLR, deep learning reconstruction; STIR, short inversion time inversion recovery; T2WI, T2-weighted imaging.

Discussion

Our results constitute the first demonstration of quantitative or qualitative improvements for image quality of routine MR sequences for morphological T-category diagnosis for NSCLC patients. Moreover, STIR imaging and CE thin-section Quick 3D imaging with CS were shown to be more sensitive and accurate for distinguishing T3 or T4 tumors from T1 or T2 tumors than unenhanced-thick- and thin-section Quick 3D imaging. In addition, STIR imaging, CE-thick-section Quick 3D imaging, and CE-thin-section Quick 3D imaging with CS were significantly more accurate for T-category assessment of NSCLC than thin-section MDCT, T2WI, and unenhanced thick-section Quick 3D imaging and unenhanced thin-section Quick 3D imaging with CS, whether DLR is applied or not. Furthermore, whether DLR is applied or not, STIR imaging and CE-thin-section Quick 3D imaging with CS were significantly more accurate for distinguishing T4- from T3-categories in NSCLC than unenhanced thick- and thin-section Quick 3D imaging with CS. This report is thus the first to demonstrate the capability of DLR and the most appropriate routine MR protocol for T-category evaluation of NSCLC.

Interobserver agreements for each qualitative image evaluation, assessment of differentiation capability for T3 or T4 cases from T1 or T2 cases, and T-category evaluation were rated substantial or almost perfect.²³ Therefore, our results are considered to be reproducible.

The application of DLR resulted in significant improvements in SNRs of tumor and muscle, overall image quality, and artifacts of all MR sequences except for unenhanced and CE-thick-section Quick 3D imaging. However, CNR between tumor and muscle did not significantly improve for any of the sequences. DLR used in this study is generated as a convolutional neural network system for denoising image noise on MR images with a low SNR and transform them into MR images with high SNR.^10–13 Our results are therefore easily speculated and considered compatible with previously reported findings.^10–13

The comparison of the diagnostic performance by all the methods for differentiating T3 or T4 from T1 or T2 tumors showed that STIR imaging and CE-thin-section Quick 3D imaging had significantly higher diagnostic performance than thin-section CT, T2WI with and without DLR, and unenhanced thick-section and thin-section Quick 3D imaging with and without DLR. In addition, the accuracy of T-category diagnosis using STIR imaging, CE-thick-section Quick 3D imaging, and CE-thin-section Quick 3D imaging were significantly higher than that using other methods. Moreover, whether DLR is applied or not, STIR imaging and CE-thin-section Quick 3D imaging with CS were significantly more accurate for distinguishing T4- from T3-categories in NSCLC than unenhanced thick- and thin-section Quick 3D imaging with CS. Therefore, STIR imaging and CE-thin-section Quick 3D imaging obtained with CS were agreed upon as the currently appropriate MR sequence for T-category assessment in routine clinical practice. Some investigators have shown that there are significant differences between malignant and benign tumors or nodes in terms of their T1 and T2 relaxation times.^24–28 Because many pathologic lesions show an increase in both T1 and T2, the addition of these two types of contrast with the STIR sequence yields a higher net tissue contrast between malignant and benign lesions.^24–28 Unlike STIR imaging, CE-thick- or thin-section Quick 3D imaging is a CE-radio-frequency spoiled 3D GRE sequence with or without CS, which allows for high-resolution imaging during a breath-hold of less than 30 s. In comparison with a conventional 3D gradient-echo sequence, the Quick 3D sequence provides similar image quality but with an increase in slice selective spatial resolution due to the use of the DFS technique. As a result, CE-thick- and thin-section Quick 3D imaging improved the CNR between tumor and normal tissues, so that the use of CS is recommended for further improvement of spatial resolution without having to prolong acquisition time in routine clinical practice. Our results are therefore compatible with the underlying MR physics for each sequence. Moreover, STIR imaging and CE-Quick 3D imaging with CS rather than that without CS merits the use in routine clinical practice.

There are several limitations to this study. First, our study evaluated 213 retrospectively selected patients on the basis of surgical or pathological results as well as consensus of our tumor board. However, there were relatively few patients in the T3 or T4 category, and this constituted a major disadvantage for the evaluation of sensitivity, specificity, and accuracy to a level beyond the decimal point. Moreover, tumor board used CT and MR images in T category, and predictive parameters for each method had some overlaps and would be better to be independent from the reference parameter. Therefore, these facts were considered as limitations in this study. Second, we utilized 320-detector row CT systems with a 64-detector row helical CT scan. However, it has been suggested that a step and shoot scan (i.e., a wide volume scan) on a 320-detector row CT system produces a higher image quality than a helical scan.^29,30 Moreover, ultra-high-resolution CT or photon counting CT, which features spatial resolution superior to that of a 320-detector row CT, was not used in this study, nor was model-based iterative reconstruction or DLR for CT.^31,32 Therefore, the diagnostic performance of thin-section MDCT in our study may have been affected by these technical factors. Third, we adopted similar diagnostic criteria for morphological assessment of mediastinal and chest wall invasions on CT and MRI, and additional criteria with specific image contrast for STIR imaging and Quick 3D imaging with and without CS. These additional criteria, based on improved image contrast or spatial resolution, are considered to be major factors for these sequences to improve their capability to distinguish T3 or T4 tumors from T1 or T2 tumors or for the accuracy of their T-category assessment. Forth, several investigators had suggested that dynamic cine MRI was useful for assessing aortic or chest wall invasion in lung cancer patients.^33–36 However, in this study, the suggested technique was not applied as a part of this study protocol. Therefore, diagnostic potential of MR imaging might be underestimated and would be able to be improved, if dynamic cine MRI is included in the MR protocols in this setting. We are therefore planning to conduct a large-scale prospective study in the near future to compare the diagnostic capability of T-category assessment among MR imaging with and without DLR and thin-section MDCT for NSCLC patients in order to determine their true significance.

Conclusion

In conclusion, DLR proved to be useful for image quality improvement on MR imaging for NSCLC patients. Moreover, in comparison with thin-section MDCT as well as other sequences, STIR imaging and CE-Quick 3D imaging with or without CS were validated as appropriate MR sequences for T-category imaging.