Objective assessment of segmentation models for thyroid ultrasound images

Abstract

Ultrasound features related to thyroid lesions structure, shape, volume, and margins are considered to determine cancer risk. Automatic segmentation of the thyroid lesion would allow the sonographic features to be estimated. On the basis of clinical ultrasonography B-mode scans, a multi-output CNN-based semantic segmentation is used to separate thyroid nodules' cystic & solid components. Semantic segmentation is an automatic technique that labels the ultrasound (US) pixels with an appropriate class or pixel category, i.e., belongs to a lesion or background. In the present study, encoder-decoder-based semantic segmentation models i.e. SegNet using VGG16, UNet, and Hybrid-UNet implemented for segmentation of thyroid US images. For this work, 820 thyroid US images are collected from the DDTI and ultrasoundcases.info (USC) datasets. These segmentation models were trained using a transfer learning approach with original and despeckled thyroid US images. The performance of segmentation models is evaluated by analyzing the overlap region between the true contour lesion marked by the radiologist and the lesion retrieved by the segmentation model. The mean intersection of union (mIoU), mean dice coefficient (mDC) metrics, TPR, TNR, FPR, and FNR metrics are used to measure performance. Based on the exhaustive experiments and performance evaluation parameters it is observed that the proposed Hybrid-UNet segmentation model segments thyroid nodules and cystic components effectively.

Introduction

The thyroid is located lower behind the front neck, a small gland in the shape of a butterfly. It regulates human metabolism. It releases a secret hormone that controls human activities, including energy, heat, heart rate, temperature, and oxygen [1, 2]. The human body is harmed when the release of hormones is improper due to the nodule's aberrant growth [1, 2]. Various imaging modalities like mammogram, Ultrasound (US), CT Scan, thermal images, and MRI have been widely used to detect thyroid nodules at an initial stage so that the patient's chance of survival can be increased [3–5]. The limitation of mammogram, MRI, CT, and thermal imaging lie in ionizing, cost, and availability, which might be harmful or easy to use to the patient [3, 4, 6]. US imaging modality is most popular to identify thyroid nodules as an initial screening test over MRI, CT, and thermal imaging [7, 8]. Owing to the disadvantages of the other imaging modalities, the US is considered as a first-line treatment to identify thyroid nodules due to its cost, availability, and no harm to the patient [9]. Low contrast and speckle noise often degrade US images' visual quality, which affect the radiologists' interpretation [10, 11]. The presence of speckle-noise reduces contrast resolution, making it more difficult to identify lesions during diagnosis [12, 13]. Over the past two decades, computer-based research of thyroid tumor US images has been extensively examined. The main objective of this work is to improve detectability during thyroid nodule screening using computer-based algorithms to assist radiologists in abnormality diagnosis [14, 15].

Segmenting thyroid US images may be difficult due to the lack of contrast between different anatomies and the existence of speckle noise [16]. Researchers have proposed different thyroid nodule segmentation techniques to more precisely segment the thyroid nodules according to their size, shape, and location [17]. Most of the segmentation methods are referred to as manual masks to indicate the segmentation algorithms. These algorithms were not suitable for real-time application for diagnosing thyroid nodules. Deep learning (DL)-based algorithms improve the graphics and do not require an initial mask. It is fully automated with an interaction time of just a few seconds and enables real-time implementation. DL-based segmentation algorithms are separated into two categories, i.e., semantic and instance segmentation. Each image pixel belongs to a particular class in the semantic segmentation, while instance segmentation splits distinct objects that belong to the same class [18–21]. Semantic segmentation classifies each pixel with a corresponding class label, i.e., belonging to the background or lesion [19]. Semantic segmentation differs from object detection as it does not predict any bounding boxes around the lesion [22]. A common architecture for semantic segmentation is based on encoder–decoder. It contains of three structure blocks, i.e., convolutional, downsampling, and upsampling. The encoder is a pre-training convolution network, while a decoder consists of a deconvolution layer, interpolation, and upsampling layer [19, 23] 24. The purpose of downsampling is to capture semantic or context information, while upsampling is to recover spatial resolution. The encoder network extracts the feature map, while the decoder network is used to recover channels' resolution. An encoder-decoder structure of segmentation algorithms is widely used to define boundaries between the images. From the executive review of literature, it is observed that semantic segmentation models have been widely used for US images of organs like heart [8, 10], kidney [25–27], breast [28–31], liver [32–34]. In the present work, an ideal despeckling algorithm is used to smoothen the image's homogenous area while preserving the edges of the lesion boundary. Semantic segmentation-based Hybrid-UNet model is proposed to segment the thyroid lesion from the original and pre-processed thyroid US images.

This paper is further arranged as: in Sect. 2 related work and significant findings represented, Sect. 3 elaborates materials and methods, in Sect. 4 result and discussions are presented and in Sect. 5 conclusion of the study is presented.

Related work and significant findings

A brief review of literature for thyroid nodules segmentation using encoder-decoder-based semantic segmentation models is summarized in Table 1.

Table 1.

A brief literature review of thyroid nodules segmentation using semantic segmentation

Investigator(s)	No. of image(s)	Pre-processing algorithm	Segmentation algorithm	Evaluation metric(s)
W. Song et al. (2015) [35]	4309	–	VGG16	mAP—98.2%
H. Ravishankar et al. (2016) [36]	140	–	Hybrid-CNN	DC—0.9
J. Ma et al. (2017) [37]	22,123	–	Self-Design	DR—0.92
Jinlian Ma et al. (2018) [38]	22,123	–	Self-Design	DR—0.95
Xuewei Li et al. (2018) [39]	300	–	FCN-TN	IoU—91%
J. Wang et al. (2018) [40]	3459	–	VGG16	IoU—0.75
S. Zhou et al. (2018) [41]	893	–	MG-UNet	DSC—0.94
X. Ying et al. (2018) [42]	1000	–	SegNet (VGG19)	IoU—87%
P. Poude et al. (2019) [43]	675 and 1600	–	U-Net	DSC—0.87 & 0.86
J. Ding et al. (2019) [44]	1936	–	ReAgU-Net	mIoU—0.78 DSC—0.86
V. Kumar et al. (2020) [45]	914	–	MPCNN	DSC—0.62
Webb, Jeremy M. (2020) [46]	120Patients	–	DeepLabv3 +	IoU—0.73
Prabal Poude et al. (2018) [47]	1416	HE & MF	U-Net	DC—0.87
Jianguo Sun et al. (2018) [48]	173	AMF & HE	FCN-AlexNet	IoU- 0.81
M. Buda et al. (2019) [49]	1278	Contrast stretching	U-Net	DCS- 0.93
Zihao Guo et al. (2020)[50]	1400	HE	DeepLabv3 +	DSC—94.08%
Gomes Ataide (2021)[51]	6066	Resizing and cropping	ResUNet	DC—0.85 & IoU—0.767

Note - HE - histogram equalization, MF - median filter, AMF - adaptive median filter, DCS/DC - dice coefficient, IoU - Intersection of Union, DR - Dice ratio, mAP - mean average precision

From the literature it is observed that different detection methods such as SSD, R-CNN, and YOLO[52–61] are commonly used to detect the bounding box around the lesion (height, width, and position) in thyroid US images. So, it is concluded that thyroid nodules instance detection does not help in the further analysis to design a CAD (computer-based diagnosis system) for characterization of thyroid nodules.

Table 1 shows that several segmentation methods based on self-design, FCN, SegNet and U-Net are used for original thyroid US images. It is observed that Zhou et al. [41] achieved the highest dice coefficient (0.94) using MG-UNet-based algorithm. In literature, very few pre-processing algorithms (only Contrast stretching, adaptive median, median filters, and Histogram equalization) used before the segmentation algorithms for pre-processing of thyroid US images. Most researchers used U-Net (DAG) architecture using simple convolution to segment thyroid nodules using pre-processed thyroid US images. Further it is also observed that limited work is available in the literature on pre-processing thyroid US images for segmentation of thyroid nodule. So, in this work efficient despeckling filters used for enhancing the performance of the segmentation model [62, 63].

Materials and methods

Here, extensive experiments conducted by combining the two publicly accessible benchmark datasets, i.e., DDTI [64] and ultrasoundcases.info (USC) [65]. Out of these two datasets 820 thyroid tumor images selected for experiment purpose. Best despeckle filtering algorithm is selected from a wide range of 64 despeckling algorithms based on diagnostically important features like structure, edge, and margin preservation. The best performing filter is selected objectively from these despeckling filters. The semantic segmentation networks are considered from a void variety of segmentation models, including SegNet (VGG16), U-Net, and proposed Hybrid-UNet. The mean intersection of union (mIoU), mean dice coefficient (mDC), TPR, TNR, FPR, and FNR metrics [66–68] used extensively for the objective assessment of segmentation models.

Workflow adopted for the segmentation of thyroid US is presented in Fig. 1. The description of phases used in this work is given below:

Fig. 1.

The workflow adopted for the segmentation of thyroid tumor US images

Phase I: dataset preparation module

Following steps involved in dataset preparation module—(a) Benchmark dataset for thyroid US images (b) Image resizing module for resizing of thyroid US image (c) True mask generation using createMask function and (d) Data/image augmentation.

Benchmark dataset for thyroid US images

820 thyroid US images selected from benchmark datasets of DDTI and ultrasoundcases.info (USC). out of these 820 images 620 images used for training. In 620 training images 36 benign thyroid tumor US images (TTUS) and 322 malignant TTUS images taken from DDTI dataset and 64 benign TTUS and 198 malignant TTUS images taken from USC dataset. For testing purpose 200 images selected from both the datasets. In 200 testing images, 30 benign TTUS and 60 malignant TTUS images taken from DDTI dataset and 70 benign TTUS and 40 malignant TTUS images taken from USC dataset.

Image resizing module

In this module, the unwanted information is removed from the thyroid US images and the resulting US images are resized. The size of the images selected is of 256 × 256 pixels. It is clinically significant to resize the images while maintaining the aspect ratio. Xiaofeng Qi et al. [69] suggested that direct resizing of the images without considering the tumor's shape or lesions and aspect ratio changes is not desirable. In order to preserve the tumor shape by maintaining the aspect ratio, adequate attention is used during resizing the thyroid US images in this work.

True mask generation

In this module, A binary mask defines image pixels belonging to the tumor or background region. The mask’s pixel values outside the lesion is set to 0, and pixel values inside the lesion is set to 1. The binary mask size is the same as that of the input image. The createMask (h) MATLAB function generates a binary mask where h defines the tumor region (ROI object) in this study. There are four different ROI objects available to create the mask h function, i.e. ellipse, point, poly, and rectangle. In the present work, the Poly object is used for generating a binary mask by expanding the interactive polygon to match the shape and size of the tumor region.

Data/image augmentation

Data/image augmentation is a technique that enlarges the dataset by applying certain transformations. 820 thyroid ultrasound images are inadequate to train a DL-based semantic segmentation network. In the present work, rotation (90° and 180°), translation, horizontal and vertical flip, and rotation of flipped images done to enlarge the 620 TTUS images which are selected for training purpose. By applying augmentation methods, 620 TTUS images converted into 11,176 images for training purpose as shown in Fig. 1.

Phase II: Despeckle filtering module

Low-contrast and speckle noise differential diagnosis between these characteristics are complex even for an experienced radiologist [9, 12]. Therefore, controlled despeckling is preferred to preserve the diagnosis information in the images, and controlled despeckling is allowed by eliminating speckle noise from homogenous areas of the region and preserving the edges of the lesion boundary. As proven in the authors' previous study DsF_FBiF and DsF_EPSF filters outperformed to reduce the speckle noise from thyroid US images by preserving edges, structure, and margin[9, 70]. So DsF_FBIF and DsF_EPSF filters used here for despeckling purpose.

Phase III: segmentation module

The main goal of segmentation is to divide an image into several distinct, non-overlapping sections that fully characterize the test object in the image [71–74]. A key task in the care of the subject being examined is the blueprint or dividing of a thyroid nodule from the background [73, 75, 76]. Generally, low SNR in ultrasound images in the form of speckle-noise degrades the margin detection or region obtained from the appropriate method [71]. The segmentation model's performance might be enhanced by despeckling US images. In the case of thyroid ultrasound, several studies have reported using conventional-based segmentation methods [44, 54, 64]. It is observed from the literature that the conventional approach becomes a semi-automatic approach for initializing the mask. Thus, an automatic approach is required. Here, in this study Hybrid-UNet semantic segmentation technique is proposed for automatic mask initialization. Hybrid-UNet is created using SegNet and UNet based semantic segmentation architecture.

Semantic segmentation

The objective of semantic image segmentation is to assign a class to each pixel in an image that corresponds to the object being represented [84–86]. The architecture of semantic segmentation based on encoder-decoder where trainable engine is called an encoder network, and pixel-wise classification is called a decoder network [41, 47, 87].

The SegNet architecture is a deep encoder-decoder model developed by a computer vision research community at Cambridge University for pixel-wise segmentation [19, 88]. It uses VGG16 pre-trained model with a convolution filter of the size of 3 × 3, batch normalization, activation-ReLU (non-linear), 2 × 2 max-pooling, and subsampling on the encoder side. The SegNet architecture is the pre-trained end-to-end network that preserves high-frequency components and decreases the trainable parameters on the decoder side [23, 24, 89].

The U-Net architecture was designed by Ronneberger, Fischer, and Bronx [90]. The U-Net architecture has two paths (a) contraction/encoder and (b) expansion/decoder path [87]. The contraction path followed by the expansion path delivers a U-shaped network. The convolution and the max-pooling layers reduce the spatial information while preserving the feature map [87, 90].

Semantic segmentation is based on FCN or classical architecture like SegNet and U-Net is trendy segmentation method for medical images. A general U-Net architecture consisting of a complete path follows a convolution network with the max-pooling operation and continued down sample feature map. As encoder in U-net, we used a relatively similar CNN architecture of SegNet (VGG-16) that consists 16 sequential layers. The entire pooling layer feature map is transferred between the encoder and decoder sides but with two convolution layers at each stage and then concatenated to perform the convolution operation. The first convolution has 8 channels, then doubles after each max pooling operation until it reaches 512. Here, Hybrid-UNet design is proposed for segmentation by using a combination of U-Net and SegNet(VGG16) [20]. The architecture of Hybrid-UNet applied in this study is presented in Fig. 2.

Fig. 2.

Hybrid-UNet architecture

The SegNet, U-Net and Hybrid-UNet models are trained via Adam optimizer with 32 mini-batch sizes and learning rates ϵ {10^–3, 10^–4, 10^–5, 10^–6}. The mini-batch size and learning rates are carefully selected so that number of images split in training so that the complete training dataset is passed in the models during training epochs and data cannot be discarded. Segmentation models trained for 30 epochs, and overfitting is avoided using early stopping criteria. These segmentation models (SegNet, U-Net, and Hybrid-UNet) are implemented on NVIDIA 1070Ti GPU with 2,432 CUDA cores.

Phase IV: Performance evaluation of proposed segmentation module

For the performance evaluation of the semantic segmentation methods researchers used different evaluation metrics like accuracy [37, 91], Precision [92, 93], Dice coefficient [73, 94], Jaccard index [95, 96], TPR [97], TNR [97], FPR [98], FNR [97], F-measure [99], Hausdorff distance [73], average distance [100], Mahalanobis distance [101] mutual information and variation of information [102]. Here the performance of SegNet, UNet and Hybrid-UNet segmentation models is calculated using the overlay region between the lesion marked by experienced participating radiologists and the lesion obtained from the segmentation model [68].

In this work, the area of interaction between ground truth (S_R) and predicted mask (S_A) calculated using the mIoU, mDC, TPR, FPR, TNR, and FNR [69–70]. The performance evaluation parameters used for the assessment of proposed segmentation algorithms are given as:

$$\mathrm{IoU}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}\mathrm{or}\frac{S_{\mathrm{A}}\cap S_{\mathrm{R}}}{S_{\mathrm{A}}\cup S_{\mathrm{R}}},$$ 1

$$\mathrm{DC}=\frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}\mathrm{or}\frac{2\left|S_{\mathrm{A}},\cap ,S_{\mathrm{R}}\right|}{\left|S_{\mathrm{A}}\right|+\left|S_{\mathrm{R}}\right|},$$ 2

$$\mathrm{TPR}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}},$$ 3

$$\mathrm{TNR}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}},$$ 4

$$\mathrm{FPR}=1-\mathrm{TNR},$$ 5

$$\mathrm{FNR}=1-\mathrm{TPR}.$$ 6

The sample confusion matrix obtained from the proposed Hybrid-UNet segmentation method is shown in Fig. 3.

Fig. 3.

Sample confusion matrix

Result and discussions

The experiments utilized to evaluate the segmentation models are listed below in Table 2.

Table 2.

Experiments conducted for the assessment of segmentation models

Experiment 1	Objective assessment of segmentation models using original thyroid US images
Experiment 2	Objective assessment of segmentation models using pre-processed thyroid US images using DsF_FBiF
Experiment 3	Objective assessment of segmentation models using pre-processed thyroid US images using DsF_EPSF

The performance of SegNet, UNet and proposed Hybrid-UNet segmentation models using original and despeckled thyroid US Images by DsF_FBiF and DsF_EPSF filters is tabulated in Table 3, and the best outcome is marked in grey.

Table 3.

The performance of semantic segmentation models using original and despeckled thyroid US Images by DsF_FBiF and DsF_EPSF Filters

Experiment (Image Type)	P.E.P	mIoU (%)	mIoU (B) (%)	mIoU (M) (%)	mDC (%)	DC (B) (%)	DC (M) (%)	TPR (%)	FNR (%)	TNR (%)	FPR (%)
	Seg. Algo
Experiment 1 (OTTUS)	Seg Net	78.4 ± 16.1	80.3 ± 15.0	77.4 ± 14.1	87.2 ± 06.4	87.3 ± 07.4	86.4 ± 06.1	85.2 ± 05.8	14.8 ± 03.8	87.6 ± 4.9	10.4 ± 3.4
	UNet	80.9 ± 11.2	81.4 ± 09.2	80.6 ± 11.1	89.1 ± 05.4	88.1 ± 04.2	87.4 ± 05.3	86.9 ± 06.9	13.1 ± 03.7	89.9 ± 3.1	10.1 ± 2.9
	Hybrid-UNet	83.1 ± 07.8	83.4 ± 08.7	82.2 ± 06.9	90.1 ± 02.9	91.4 ± 03.2	89.4 ± 0.02	88.2 ± 04.9	11.8 ± 02.8	90.8 ± 3.2	09.2 ± 2.3
Experiment 2 (DTTUS by DsF_FBiF Filter)	SegNet	79.4 ± 13.2	80.3 ± 11.1	79.1 ± 12.4	88.7 ± 05.8	89.4 ± 06.2	88.1 ± 04.8	87.6 ± 04.9	12.4 ± 03.1	88.2 ± 4.1	08.8 ± 2.9
	UNet	82.1 ± 09.7	84.0 ± 07.1	82.8 ± 09.2	89.3 ± 05.1	89.9 ± 05.7	88.8 ± 04.3	87.9 ± 05.9	12.1 ± 03.8	89.1 ± 3.1	10.9 ± 2.5
	Hybrid-UNet	85.2 ± 06.2	86.2 ± 05.8	84.4 ± 05.7	90.6 ± 02.8	90.7 ± 0.02	89.4 ± 0.03	89.7 ± 04.1	10.3 ± 02.7	90.8 ± 3.2	09.2 ± 2.1
Experiment 3 (DTTUS by DsF_EPSF Filter)	SegNet	80.4 ± 12.8	81.4 ± 11.7	79.3 ± 15.4	90.0 ± 05.1	90.4 ± 04.9	89.9 ± 05.2	89.1 ± 04.9	10.9 ± 03.0	90.4 ± 4.1	09.6 ± 3.1
	UNet	83.2 ± 09.7	85.1 ± 08.5	83.6 ± 11.1	92.9 ± 03.9	93.1 ± 03.8	91.9 ± 04.1	92.3 ± 04.9	07.7 ± 02.8	93.2 ± 3.3	06.8 ± 2.9
	Hybrid-UNet	86.6 ± 09.8	88.8 ± 8.7	85.6 ± 09.5	93.2 ± 03.1	93.9 ± 02.8	92.8 ± 03.2	90.5 ± 03.1	09.5 ± 02.6	94.9 ± 2.7	05.1 ± 2.0

Note - P.E.P. - performance evaluation parameters, OTTUS - Original Thyroid Tumor US Images, DTTUS - Despeckled Thyroid Tumor US Images, B - benign, M - malignant

In Table 3 the performance of semantic segmentation models (i.e. SegNet, U-Net and Hybrid-UNet using original and despeckled thyroid US images is presented in terms of mIoU, mDC, TPR, TNR, FPR, and FNR. It is noticed that proposed Hybrid-UNet segmentation model perform better segmentation, with the highest values of mIoU (86.6%) & mDC (93.2%) for thyroid US images filtered by DsF_EPSF while separately in case of benign tumor images 88.8% mIoU 93.9% mDC and for malignant tumor images 85.6% mIoU & 92.8% mDC noticed. As shown in Table 4, the impact of the despeckling algorithm on the performance of the proposed Hybrid-UNet segmentation model is also examined.

Table 4.

Effect of despeckling algorithm on the performance of proposed Hybrid-UNet segmentation model

Type of Images	Assessment Metric for Despeckle filter (SEPI)	Overall mIoU (%)	Overall mDC (%)	TPR (%)	FNR (%)	TNR (%)	FPR (%)
Original	–	83.1 ± 7.8	90.1 ± 2.9	88.2 ± 4.9	11.8 ± 2.8	90.8 ± 3.2	09.2 ± 2.3
DsF_FBiF	0.98 ± 0.02	85.2 ± 6.2	90.6 ± 2.8	89.7 ± 4.1	10.3 ± 2.7	90.8 ± 3.2	09.2 ± 2.1
DsF_EPSF	0.99 ± 0.01	86.6 ± 9.8	93.2 ± 3.1	90.5 ± 3.1	9.5 ± 2.6	94.9 ± 2.7	05.1 ± 2.0

In Fig. 4, the sample thyroid tumor US original image(a₁), despeckled image by DsF_FBiF(b₁) and DsF_EPSF(c₁), and corresponding tumor marked by radiologist, true mask, predicted mask, predicted lesion and overlay diagram between true TTUS lesion and predicted TTUS lesion using SegNet, UNet and Hybrid-UNet models are presented.

Fig. 4.

a₁ Original Thyroid Tumor Ultrasound (TTUS) image, b₁ despeckled TTUS image by DsF_FBiF, c₁ Despeckled TTUS image by DsF_EPSF filter, a₂–c₂ original TTUS images with tumor marked by radiologist, a₃–c₃ True TTUS masked images, a₄–a₆ predicted masks (applied on Original TTUS) images using SegNet, UNet and Hybrid-UNet, b₄–b₆ predicted masked (applied on despeckled TTUS by DsF_FBiF filter) images using SegNet, UNet and Hybrid-UNet, c₄–c₆ predicted masked (applied on despeckled TTUS by DsF_EPSF filter) images using SegNet, UNet and Hybrid-UNet, a₇–a₉ predicted lesions (applied on Original TTUS) images using SegNet, UNet and Hybrid-UNet, b₇–b₉ predicted lesions (applied on despeckled TTUS by DsF_FBiF filter) images using SegNet, UNet and Hybrid-UNet, c₇–c₉ predicted lesions (applied on despeckled TTUS by DsF_EPSF filter) images using SegNet, UNet and Hybrid-UNet, a₁₀–a₁₂ overlay diagram between true TTUS lesion and predicted TTUS lesion (applied on Original TTUS) images using SegNet, UNet and Hybrid-UNet, b₁₀–b₁₂ overlay diagram between true TTUS lesion and predicted TTUS lesion (applied on despeckled TTUS by DsF_FBiF filter) images using SegNet, UNet and Hybrid-UNet, c₁₀–c₁₂ overlay diagram between true TTUS lesion and predicted TTUS lesion (applied on despeckled TTUS by DsF_EPSF filter) images using SegNet, UNet and Hybrid-UNet

From Table 4, it is observed that the values of performance evaluation parameters are improved using despeckled thyroid US images. The result obtained from the objective assessment of segmentation models using original and despeckled thyroid US images indicates that the DsF_EPSF filter yield better segmentation results in terms of mIoU, mDC, TPR, TNR, FPR, and FNR by using proposed Hybrid-UNet segmentation model.

A comparison of the proposed methodology Hybrid-UNet segmentation model with other existing techniques for thyroid tumor segmentation is tabulated in Table 5.

Table 5.

Comparison of the proposed methodology (Hybrid-UNet segmentation model with other existing techniques for thyroid tumor segmentation

Investigator (s)	Number of images	Technique used for pre-processing	Segmentation model(s)	Evaluation metrics
Case 1: Using Original TTUS Images
S. Zhou et al. [41]	893	–	MG-U-Net	Dice Coefficient -0.9
J. Ma et al. [37]	10,357	–	Self-design	Dice ratio- 0.9
Proposed method (Hybrid_UNet)	820	–	Hybrid-UNet	mIoU—83.1 ± 7.8 and mDC = 90.1 ± 2.9
Case 2: Using pre-processed TTUS Images
Prabal Poude et al. [47]	416	Median Filter and Histogram Equalization	UNet	Dice Coefficient = 0.876
J. Sun et al. [48]	173	Adaptive Median Filter and Histogram Equalization	FCN-AlexNet	mIoU = 0.81
M. Buda et al. (2019) [49]	1278	Contrast stretching	U-Net	Dice Coefficient = 0.93
Zihao Guo et al. (2020)[50]	1400	Histogram Equalization	DeepLabv3 +	Dice Coefficient = 0.94
Gomes Ataide et al. (2021)[51]	6066	Resizing and cropping	ResUNet	mDC—0.857 & mIoU—0.767
Proposed method (Hybrid_UNet)	820	DsF_EPSF	Hybrid-UNet	mIoU—86.6 ± 9.8 and mDC = 93.2 ± 3.1

From Table 5 it is concluded that, Hybrid-UNet performs better in comparison to other techniques when pre-processed TTUS images using DsF_EPSF are used as input to the proposed Hybrid-UNet segmentation model.

Conclusion

In this study, extensive experiments performed to segment the thyroid tumor US images using SegNet, UNet and Hybrid-UNet using original and despeckled thyroid tumor US images using DsF_EPSF and DsF_FBiF filters. The performance of Hybrid-UNet segmentation model is compared with existing segmentation models (i.e. SegNet and UNet) in terms of mIoU, mDC, TPR, TNR, FPR, and FNR metrics. From the Tables 3 and 4 it is noticed that the Hybrid-UNet segmentation method yields better segmentation in terms of shape, margin, composition, and echogenic characteristics exhibited by lesions. In case of segmentation of original thyroid US images using proposed segmentation model 83.1% mIoU and 90.1% mDC achieved while 86.6% mIoU and 93.2% mDC achieved when despeckled TTUS images by DsF_EPSF filter are given as input to the proposed method.

The proposed model can be used to make things simpler i.e. extracting the thyroid organ with region of interest and transform the thyroid US images into meaningful subject. An efficient CAD tool for analysis and classification of thyroid US Images can be designed to take as a second opinion during the clinical treatment. It is concluded that Hybrid-UNet segmentation using DsF_EPSF filtered images, yields more clinically acceptable enhanced segmented diagnostic information concerning the proper shape, size, and margins of thyroid tumor.

Funding

This study and authors not received any funding from other sources.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.