Focal therapy of prostate cancer: Use of artificial intelligence to define tumour volume and predict treatment outcomes

Abstract

Objectives

The aim of this study is to evaluate new software (Unfold AI) in the estimation of prostate tumour volume (TV) and prediction of focal therapy outcomes.

Subjects/patients and methods

Subjects were 204 men with prostate cancer (PCa) of grade groups 2–4 (GG ≥ 2), who were enrolled in a trial of partial gland cryoablation (PGA) at UCLA from 2017 to 2022. Magnetic resonance imaging (MRI)‐guided biopsy (MRGB) was performed at diagnosis and at 6 and 18 months following PGA. Utilising Unfold AI (FDA‐cleared 2022), which generates a 3D map of GG ≥ 2 PCa margins, we retrospectively estimated TV for each patient. TV was compared against conventional baseline variables as a correlate of a successful primary outcome—defined here as the absence of GG ≥ 2 on follow‐up MRGB at 6 months. Secondary outcomes were MRGB at 18 months and failure‐free survival, that is, lack of metastasis or salvage whole gland therapy. Receiver operating curves and multivariate analysis were used to determine significance.

Results

A successful primary outcome was observed in 77.7% of patients. Significant correlates of a successful ablation were percent pattern 4 and TV; areas under the curve (AUCs) were 0.60 and 0.73, respectively. GG was not a correlate of success (AUC = 0.51). A TV of 1.5 cc provided the optimal combination of sensitivity (55.8%) and specificity (85.7%) at 6 months. TV was also significantly associated with secondary outcomes. In multivariate analysis, TV was the variable most associated with 6‐ and 18‐month biopsy success (adjusted odds ratios [aORs] were 6.1 and 4.2). Utilising TV ≤ 1.5 cc as a PGA criterion would have prevented 72% of failures at the cost of 42% of successes.

Conclusion

The AI‐based software Unfold AI estimates TV, which is significantly associated with biopsy outcomes after focal cryoablation. The rate of treatment success is inversely related to TV.

1. INTRODUCTION

The introduction of multiparametric magnetic resonance imaging (MRI) for prostate cancer (PCa) catalysed the adoption of partial gland ablation (PGA), which has become an option for men with localised disease. ¹ , ² Although MRI localises most PCa, it underestimates tumour volume (TV) and may miss small satellite tumours. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ This underestimation of TV may be partially responsible for the high recurrence rates after PGA. ¹ , ⁹ , ¹⁰ In prior trials with close follow‐up, PGA has led to a success rate varying from 65% to 97%, depending on how success is defined. ⁹ , ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ Furthermore, the criteria for identifying which patients are appropriate to treat with PGA lack consensus, contributing to variable outcomes. While tumour grade and prostate‐specific antigen (PSA) are commonly cited, ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ there is little consensus regarding TV and treatment margin criteria. ⁵ , ⁸ , ²¹ Thus, many treatment failures could be prevented by improving the delineation of TV and refining criteria for PGA eligibility.

Newer imaging modalities, such as PSMA PET CT and micro‐ultrasound, provide only modest improvements over MRI and introduce new limitations, namely, low tissue resolution and lesion conspicuity. ²² , ²³ Alternatively, artificial intelligence (AI) may be used to incorporate multimodal information from imaging, PSA and targeted biopsy, helping to predict TV and margins. FDA‐cleared AI software (Unfold AI, Avenda Health) estimates tumour risk in 3D and accurately predicts intraprostatic margins compared to surgical pathology. ²⁴ Utilising AI may provide a means of determining PGA candidates based on predicted tumour boundaries and volume.

We hypothesise that TV is associated with the outcome of PGA. Herein, we evaluate the relationship of TV derived from Unfold AI (and other clinical variables) to outcomes of PGA with cryotherapy.

2. METHODS

2.1. Design

This study evaluated 204 men treated with PGA cryotherapy at UCLA Medical Center between February 2017 and December 2022. Clinical data were acquired prospectively within a single‐arm clinical trial (NCT 03503643). Before enrollment, men were counselled on the nature of PGA and offered the alternatives of surgery, radiation or active surveillance depending on tumour grade. All patients provided informed consent and initially agreed to surveillance biopsy at 6 and 18 months. We defined success as lacking International Society of Urological Pathology grade group (GG) ≥ 2 prostate cancer.

2.2. Subjects

Eligible candidates included men with unilateral GG ≥ 2 prostate cancer. Patients were diagnosed by MRI‐guided biopsy (Artemis, Eigen Corp.). Inclusion criteria specified that the area of intended ablation (index lesion + surrounding GG ≥ 2 systematic cores) was amenable to hemigland cryoablation. Men with GG1 outside the intended ablation remained eligible to enrol. The trial protocol also specified men aged 40–85, PSA < 20 mg/mL, prostate volume <70 cc and a life expectancy > 10 years. We excluded men with prior prostate cancer treatment, including androgen deprivation.

2.3. Treatment

The cryotherapy treatment (PGA) at UCLA has been previously described. ⁹ Briefly, men underwent treatment utilising ultrasound‐guided transperineal placement of 14‐gauge argon cryoprobes (Galil Medical, Inc.) under general anaesthesia. The treatment margin was determined using in‐house software displaying the MRI, regions of interest (ROIs) and biopsy cores with cancer status. ⁹ This information was cognitively transferred to real‐time ultrasound for gland ablation. Cryoablation included two freeze–thaw cycles. We utilised thermal probes and ultrasound to monitor treatment progress and temperature relative to the rectum and external sphincter. Patients were discharged on the day of surgery with a Foley catheter, which was removed after 2 days.

2.4. Outcomes

The primary outcome was the presence (failure) or absence (success) of GG ≥ 2 following cryotherapy at 6 months. Patients obtained a 3‐Tesla MRI and MRI‐guided biopsy at 6 and 18 months following cryoablation. The 6‐month biopsy was evaluated for in‐field failures by returning to previously positive core locations and adding systematic cores ipsilateral to the ablation. At the 18‐month biopsy, which was considered a secondary outcome, the 6‐month biopsy protocol was repeated and systematic cores in the nonablated hemisphere were also collected. Any new MRI targets were biopsied at either time point. The biopsy technique and protocol following cryoablation are detailed in prior publications. ⁹ , ²⁵ , ²⁶ , ²⁷ For this study, patients failing at 6 months were not included in the analysis at 18 months. Patients with biopsy failure could proceed to active surveillance, salvage PGA or salvage whole gland therapy (whole gland ablation, surgery or radiation), depending on the extent of recurrence. Failure‐free survival (FFS), a secondary endpoint, was defined as lacking whole gland treatment, N1 or M1 cancer. ¹⁰ , ¹³ , ¹⁴ , ¹⁵ , ²⁸ Within our cohort, FFS data were consistently available for 2 years following treatment.

2.5. AI TV calculation

AI TVs were retrospectively calculated by co‐author A.P., using AI software (Unfold AI) previously validated to improve focal therapy tumour margins. ²⁴ , ²⁹ This software generates a patient‐specific cancer map, wherein the risk of prostate cancer is estimated at each voxel (3D pixel) within an MRI (Figure 1). Algorithm inputs include PSA, T2‐weighted axial MRI and each biopsy core's GG, cancer length and 3D location. The Unfold AI default output is a boundary that encapsulates the maximal extent of tumour, including a margin of normal tissue (Sensitivity 97%, Figure 1: moderate‐risk boundary). However, users can reduce the extent of normal tissue margin through a web‐based interface. In this study, prostate TV was approximated by reducing the normal tissue margin to near‐zero, with an estimated <1% chance of negative PGA margins (Sensitivity 65%, Figure 1: high‐risk boundary). A supporting information video on TV calculation is available online (https://youtu.be/Et43yXueS0Q). To evaluate our approach of identifying TV, we compared against 97 radical prostatectomy specimens and the histopathology volume of tumours 3D‐reconstructed from pathologist slide annotations (R ² = 0.76; Figure S1). By comparison, the Prostate Imaging Reporting and Data System (PI‐RADS) ROI volume had a poor correlation with histopathology volume (R ² = 0.33, p < 0.01). TV was available for 158 patients within the PGA cohort. We wish to note that a consideration of TV is the nonspherical nature of prostate tumours. A tumour with a large maximal dimension (for example, 2.6 cm) quickly dwindles in size due to the ‘pancaking’ effect of the transition zone on the peripheral zone (Figure S2). This prostatectomy cohort was not used in the development of the AI cancer mapping algorithm but for the correlation of maps with TV.

FIGURE 1.

Representative illustration of (A) AI tumour volume at different risk boundaries compared with (B) whole mount surgical pathology of a patient who was assessed for PGA but received radical prostatectomy. The high‐risk boundary (red) includes voxels with a 50% positive predictive value and 65% sensitivity. The moderate‐risk boundary, by contrast, has a 97% sensitivity but encompasses larger volumes of benign tissue. For the purposes of this study, the high‐risk boundary was considered the ‘AI tumour volume.’ Note that the MRI‐ROI (grey outline) is far smaller than the AI tumour volume. A supporting information video on TV calculation using the unfold AI software interface is available online. AI, artificial intelligence; MRI, magnetic resonance imaging; PGA, partial gland ablation; ROI, region of interest.

2.6. Statistical analysis

We evaluated the ability of baseline patient variables to predict primary and secondary outcomes. Given that many consensus statements utilise the biopsy GG to determine PGA eligibility, we compared baseline variables against the predictive accuracy of the GG. ²¹ We first stratified variables according to GG and evaluated differences using Chi‐square and Kruskal–Wallis tests. We assessed predictive accuracy using receiver operator characteristic (ROC) analysis and compared area under the curve (AUC) measures utilising DeLong's tests.

For baseline variables found to be significantly better than GG with AUC 95% confidence intervals (CI) above 0.5 at 6‐month biopsy, we determined the optimal sensitivity and specificity (balanced accuracy). We calculated changes in PGA outcomes as if these baseline variables had been considered exclusion criteria. We calculated the TV threshold for PGA, which maximised the successful treatments and minimised failures.

Finally, we performed a multivariate analysis to evaluate the relative impact of baseline variables on biopsy success and FFS. Adjusted odds ratios (aORs) were calculated. Given that AI utilises multiple input variables, we calculated the variance inflation factor to assess whether AI TV was collinear with other variables. All analyses were conducted in SAS 9.4 (SAS Institute).

3. RESULTS

In Table 1, the patient characteristics of all 204 men are stratified by baseline GG. As the GG increased, men were older, but ROI diameter and CCL were smaller. Otherwise, differences between groups were not significant.

TABLE 1.

Patient characteristics categorised by grade group.

		Grade group at baseline
	Total (N = 204)	GG 2 (N = 135)	GG 3 (N = 56)	GG 4 (N = 13)	P‐value
Race/ethnicity, n (%)					0.2148 a
Asian/Pacific Islander	7 (3.4%)	4 (3.0%)	2 (3.6%)	1 (7.7%)
Black/African American	8 (3.9%)	6 (4.4%)	1 (1.8%)	1 (7.7%)
Hispanic/Latinx	3 (1.5%)	3 (2.2%)	0 (0.0%)	0 (0.0%)
White	126 (61.8%)	75 (55.6%)	43 (76.8%)	8 (61.5%)
Other	12 (5.9%)	7 (5.2%)	4 (7.1%)	1 (7.7%)
Unknown	48 (23.5%)	40 (29.6%)	6 (10.7%)	2 (15.4%)
Age at treatment					0.0011 b
Median (IQR)	69.0 (64.0, 74.0)	67.0 (63.0, 72.0)	71.0 (65.0, 76.0)	75.5 (68.5, 78.0)
Missing	11	7	3	1
PSA (ng/mL)					0.1052 b
Median (IQR)	6.7 (4.8, 10.2)	6.5 (4.7, 9.5)	7.9 (5.3, 11.4)	7.2 (6.4, 10.5)
Missing	0	0	0	0
PSA density (ng/mL) b					0.9411 b
≤0.15	93 (46.0%)	63 (46.7%)	25 (45.5%)	5 (41.7%)
>0.15	109 (54.0%)	72 (53.3%)	30 (54.5%)	7 (58.3%)
Missing	2	0	1	1
% Free PSA					0.7461 b
Median (IQR)	14.0 (10.0, 18.0)	14.0 (11.0, 18.0)	14.0 (9.0, 21.0)	12.0 (10.0, 17.0)
Range	5.0, 68.0	6.0, 68.0	5.0, 40.0	8.0 22.0
Missing	51	29	18	4
Prostate volume (mL)					0.1123 b
Median (IQR)	43.0 (33.0, 55.5)	40.0 (30.8, 53.7)	49.0 (35.0, 56.0)	50.5 (38.0, 58.5)
Missing	12	7	4	1
PI‐RADS, n (%)					0.9566 a
2	21 (10.3%)	13 (9.6%)	7 (12.7%)	1 (7.7%)
3	24 (11.8%)	17 (12.6%)	6 (10.9%)	1 (7.7%)
4	84 (41.4%)	58 (43.0%)	20 (36.4%)	6 (46.2%)
5	74 (36.5%)	47 (34.8%)	22 (40.0%)	5 (38.5%)
Missing	1	0	1	0
MRI ROI diameter					<0.0001 b
Median (IQR)	14.0 (10.0, 17.0)	14.0 (11.0, 18.0)	14.5 (11.0, 18.0)	5.0 (2.0, 7.0)
Missing	22	12	8	2
Cancer core length c					<0.0001 b
Median (IQR)	7.0 (5.0, 9.0)	8.0 (6.0, 10.0)	6.0 (4.0, 8.0)	5.0 (2.5, 7.0)
Missing	3	3	0	0
% Pattern 4					<0.0001 b
Median (IQR)	30.0 (20.0, 60.0)	20.0 (10.0, 30.0)	60.0 (60.0, 70.0)	100.0 (100.0, 100.0)
Missing	5	2	3	0

Abbreviations: GG, grade group; MRI, magnetic resonance imaging; PSA, prostate‐specific antigen; ROI, region of interest.

^a Chi‐square p‐value.

^b Kruskal–Wallis p‐value.

^c From the highest grade core.

In Table 2, results of partial gland cryotherapy, stratified by GG, are shown. Successful cryoablation (absence of >GG2 on biopsy) was 77.7% among all patients at 6 months, declining to 60.1% at 18 months. Among the 10 men with GG4, the success rate was 100% at 6 months. Otherwise, GG was not a significant predictor of success; only a slight advantage was found for men with GG2 versus men with GG3. Twenty‐three went on to salvage treatments, most often surgery or radiation therapy. FFS at 18 months was slightly better for men with GG2 than others.

TABLE 2.

Results of partial gland cryoablation.

		Grade group at baseline
	Total (N = 204)	GG 2 (N = 135)	GG 3 (N = 56)	GG 4 (N = 13)	P‐value
Absence of GG ≥ 2
6 months	150 (77.7%)	102 (79.1%)	38 (70.4%)	10 (100.0%)	0.0234 a
Declined biopsy	11	6	2	3
18 months	86 (60.1%)	60 (61.9%)	20 (55.6%)	6 (60.0%)	0.7709 a
Awaiting biopsy	61	38	20	3
Metastasis, n (%)	2 (0.01%)	2 (0.01%)	0 (0.00%)	0 (0.00%)	0.9999 a
Radiation, n (%)	8 (0.04%)	1 (0.01%)	6 (0.11%)	1 (0.08%)	0.0023 a
Surgery, n (%)	10 (0.05%)	5 (0.37%)	4 (0.07%)	1 (0.08%)	0.3731 a
Whole gland salvage ablation, n (%)	3 (0.01%)	2 (0.01%)	1 (0.02%)	0 (0.00%)	0.9999 a
Failure‐free survival, n (%)	181 (88.73%)	124 (91.85%)	46 (82.14%)	11 (84.62)	0.1378 b

Abbreviation: GG, grade group.

^a Fisher's exact p‐value.

^b Chi‐square p‐value.

In Figure 2, ROC curves are shown for three baseline metrics of possible relevance in the prediction of treatment outcomes. At the 6‐month biopsy, % pattern 4 and AI TV were both superior to GG and had a 95% CI above 0.5 (Figure 2). As shown in the figure, the AUC for TV (0.72) was significantly better than GG, PSAD and % pattern 4. All other variables were not significant (Table S1). We analysed patients with Unfold AI regarding the position of the ROI and successful treatment. We stratified the ROI position by anterior/posterior, apical/basal and medial/lateral. Stratification by any position metric did not significantly impact success at any endpoint. We also evaluated success by TV; we found that success was inversely related to tumour size (Figure 3).

FIGURE 2.

ROC curves for prediction of results after cryotherapy PGA, showing performance of AI tumour volume, % pattern 4, PSA density and grade group at 6 months. Note that the AUC for tumour volume was 0.72, significantly greater than that for other variables (p < 0.01). At 18‐month biopsy and FFS, PSA density and AI tumour volume had 95% CI above 0.5 but were not significantly superior to the grade group (Table S1). AI, artificial intelligence; AUC, area under the curve; FFS, failure‐free survival; MRI, magnetic resonance imaging; PGA, partial gland ablation; PSA, prostate‐specific antigen; ROC, receiver operator characteristic.

FIGURE 3.

Proportion of patients biopsied at 6 and 18 months with successful PGA versus the tumour volume (predicted using unfold AI). AI, artificial intelligence; PGA, partial gland ablation.

We determined a threshold for % pattern 4, PSA density and AI TV to maximise balanced accuracy at 6‐month biopsy, aiming to define the optimal inclusion criteria of PGA. Thresholds for % pattern 4 were selected at 20%, PSA density at 0.15 and TV at 1.5 cc. Accuracy metrics for these values are included in Table S2. We applied the prior values as hypothetical exclusion criteria within this cohort to evaluate the impact on biopsy success. Excluding patients with % pattern 4 > 20% improved success rates at 6‐month and 18‐month biopsies by 9% (76% to >85%) and 0% (60% to >60%), respectively. Excluding patients with a PSA density >0.15 improved success rates by 3% (77% to >79%) and 6% (61% to >67%), respectively. Excluding patients with an AI TV of >1.5 cc improved success rates by 17% (76% to >93%) and 11% (56% to >67%), respectively. Combining the impact of the 6‐ and 18‐month biopsies, excluding patients with an AI TV >1.5 cc, would have prevented 72% of our cohort's failures at the cost of 42% of successes. Example cases are provided in Figures 4 and 5.

FIGURE 4.

Example cases of focal GG2 lesions with (A) a small tumour volume (TV) of 0.3 cc (0.6% of prostate volume, PV), (B) a moderate TV of 1.9 cc (4.7% of PV) and (C) a large TV of 7.4 cc (20.8% of PV). TV (derived from unfold AI) is shown in black outline. MRI‐visible lesion is shown as a grey dotted line. For the first case, treatment was successful; for the second case, residual disease was absent at 6‐month biopsy but present at 18‐month biopsy; for the third case, residual disease was detected at 6‐month biopsy. Note that the initial MRI‐visible lesion (grey) is similar for each of the three cases but much less than AI‐derived TV for the second and third case. For these cases, the results of ablation were directly related to the volume of the tumour. AI, artificial intelligence; GG, grade group; MRI, magnetic resonance imaging; TV, tumour volume.

FIGURE 5.

A case where PGA successfully (top row) and unsuccessfully (bottom row) eliminated GG ≥ 2 on follow‐up biopsy. The left columns show pretreatment data, including MRI, core locations from the pretreatment planning biopsy, the cancer estimation map and the AI tumour volume (TV) outlined in black. The right column shows posttreatment data, including MRI and combined core locations from both the 6‐ and 18‐month follow‐up biopsy. AI, artificial intelligence; GG, grade group; MRI, magnetic resonance imaging; PGA, partial gland ablation; TV, tumour volume.

While a TV of 1.5 cc provided the highest balanced accuracy, we also evaluated the TV threshold, yielding the most correct decisions for proceeding with PGA while avoiding failures. A slightly higher TV of 2.4 cc optimised correct decisions while avoiding failure. An AI TV threshold of 2.4 cc would have prevented 52% of failures at the cost of 24% of successes (Figure 6).

FIGURE 6.

(A) Plot demonstrating the impact of AI TV thresholds on sensitivity, specificity and balanced accuracy for a combined 6‐ and 18‐month biopsy outcome. The maximum balanced accuracy occurs at an AI TV threshold of 1.5 cc, though a second local maximum occurs at 2.4 cc (52% of failures prevented at the cost of 24% of successes). (B) Plot demonstrating the impact of AI TV thresholds on the number of successes, prevented failures and ‘correct decisions,’ defined as ‘true positives + true negatives − false positives − false negatives.’ The maximal number of correct decisions occurs at an AI TV threshold of 2.4 cc. AI, artificial intelligence; TV, tumour volume.

On multivariate analysis, we found that only AI TV was associated with PGA success at both 6‐ and 18‐month biopsies (aOR: 6.14 and 4.23; Table 3). Percent pattern 4 was associated with PGA success at 18 months but had an aOR of 1.03 (Table 3). None of the variables predicted FFS, and AI TV was not collinear with other variables.

TABLE 3.

Multivariate analysis of baseline covariate to predict treatment failure (GG ≥ 2 at follow‐up Bx).

	Outcome
	Grade group <2 at 6 months	Grade group <2 at 18 months	Failure‐free survival
	aOR (95% CI)	aOR (95% CI)	aOR (95% CI)
Tumour grade
Grade group (GG 2 vs. GG 3/4)	0.268 (0.038, 1.913)	4.673 (0.741, 29.474)	0.228 (0.022, 2.409)
% Pattern 4	0.970 (0.940, 1.002)	1.032 (1.001, 1.065)	0.986 (0.947, 1.026)
Tumour volume
Cancer core length	0.903 (0.787, 1.036)	1.024 (0.879, 1.192)	1.055 (0.877, 1.270)
ROI diameter	0.987 (0.930, 1.047)	1.054 (0.949, 1.172)	1.014 (0.937, 1.098)
ROI volume	0.958 (0.644, 1.425)	0.889 (0.436, 1.813)	0.959 (0.593, 1.550)
AI tumour volume	6.141 (2.077, 18.154)	4.234 (1.533, 11.691)	0.307 (0.079, 1.185)
Tumour grade/volume
PSA	0.974 (0.888, 1.069)	0.973 (0.869, 1.089)	1.056 (0.959, 1.163)
PSA density (≤0.15 vs. >0.15)	0.647 (0.239, 1.751)	1.712 (0.607, 4.828)	0.670 (0.179, 2.505)
PI‐RADS (3 vs. 4/5)	2.349 (0.458, 12.038)	1.152 (0.319, 4.157)	1.609 (0.292, 8.856)

Note: Values in bold emphasis indicate statistically significant with a 95% confidence interval greater than 1.

Abbreviations: aOR, adjusted odds ratio; IQR, interquartile range; GG, grade group; PSA, prostate‐specific antigen; ROI, region of interest.

4. DISCUSSION

In this study, 204 men with localised intermediate and high‐risk PCa underwent PGA‐cryoablation. The value of conventional variables (e.g. PSAD, GG, %Pattern 4) was compared against a new metric—AI‐derived TV—in the prediction of treatment success (i.e., no ≥ GG2 on follow‐up biopsy). All men had clinically significant PCa (including some with GG4), and all had outcomes determined by MRI‐guided biopsy. We found GG to be a poor predictor of success, whereas AI TV had substantial predictive accuracy. Other tumour characteristics—ROI diameter, ROI volume and PI‐RADS score—also performed poorly. Among men considering PGA, utilising an AI TV >1.5 cc as an exclusion criterion would have prevented 72% of failures at a cost of excluding 42% of successes. Increasing the threshold to 2.4 cm would have prevented 52% of failures at the cost of 24% of successes.

Prior consensus papers on PGA candidacy have emphasised baseline GG as an exclusion criterion. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ The present analysis suggests that TV should be considered for PGA candidacy, even more so than GG. Here TV was found to predict the outcome of PGA, using strict biopsy endpoints, comparable to other less rigorous studies where assignment of success was not always determined by biopsy. ¹³ , ¹⁴ , ¹⁵ Determination of TV may constitute a new method to predict success rates of focal therapy.

Large TV may account for the higher persistence of cancer within this cohort. An NYU cohort with a similar treatment technique and biopsy follow‐up achieved a 93% success rate at 3 years. ¹¹ A direct TV comparison between these two cohorts is not currently available; however, there is evidence that TV contributed to differences in success. A comparison of the UCLA and NYU baseline characteristics (Table 1 values) demonstrated a significant increase in number of PI‐RADS 5 lesions treated at UCLA (p < 0.01).

The present findings suggest that large TVs increase the risk of in‐field failure. However, the determination of TV using only measurements of an MRI‐visible lesion is not sufficient. MRI underestimation of TV ³ , ⁴ , ⁵ , ⁸ is an important factor leading to the high recurrence rates seen following PGA. ¹ , ⁹ , ¹⁰ Thus, wide treatment margins are required for complete tumour ablation. ⁴ , ⁵ , ⁷ , ⁸ , ²⁴ However, even with wide margins, a similarly sized ROI can represent a variety of tumour shapes (Figure 3), making uniform margins around the ROI prone to error. An AI approach to predict tumour margins integrates multiple data elements, helping to clarify not only volume but shape of the tumour. Integrating multimodal data to plan PGA margins will hopefully improve PGA successes (Figures 3 and 4).

Limitations of the study include a nonrandomised design set in a single institution, which has accrued substantial experience with MRI‐guided biopsy and focal therapy over 10–15 years. Additionally, the data indicate considerable focal therapy success with GG4 tumours. While TV was shown to be an important determinant of PGA success, tumour grade remains relevant. Several studies have identified tumour grade as a risk for PGA failure, ¹³ , ¹⁴ and high‐grade index tumours are likely induce the proliferation of satellite tumour that are difficult to identify. ³⁰ , ³¹ Notably, however, in the present study, many of the GG4 tumours were small according to ROI diameter and cancer core length.

Furthermore, we had no method to evaluate the efficacy and extent of cryotherapy. Our technique is similar to others published in the literature with two freeze–thaw cycles, temperature probe safety monitoring and ≥1 cm margins. ⁹ However, as previously reported, men treated in the second half of the study were more likely to succeed at 18 months (86% vs. 56%, p = 0.02). ⁹ Therefore, this cohort includes our institution's learning curve.

Despite these limitations, all subjects were studied prospectively using a uniform protocol mandating biopsy follow‐up. Biopsy, MRI interpretation and cryotherapy procedures were all performed by experienced personnel. A multicentre randomised trial to evaluate TV in selecting focal therapy candidates is indicated.

5. CONCLUSION

In this study, we found that AI‐derived TV had better predictive accuracy for PGA failure than the tumour GG. Within our cohort, using AI‐determined TV >1.5–2 cc as an exclusion criterion would have reduced the number of cryotherapy failures by over half.

CONFLICT OF INTEREST STATEMENT

Unfold AI, software used to determine prostate tumour margins, was cleared by the US FDA in December 2022 and is commercially available through Avenda Health (www.avendahealth.com). Dr. Brisbane receives no financial compensation from Avenda Health. He had complete control of the data and manuscript preparation. Dr. Priester serves as a Senior Data Scientist at Avenda Health. Dr. Natarajan and Dr. Marks are co‐founders of Avenda Health.

Supporting information

Figure S1: (A,C) Cancer estimation maps with estimated TV; (B,D) Ground‐truth pathology slides for a large tumour (top row) and a small tumour (bottom row); (E) scatterplot showing the correlation between AI model output (volume of voxels with high cancer risk; cancer risk > 50%) and histopathology‐proven tumour volume, R2 = 0.76.

Figure S2: Example tumour volume of 1.4 cc. Note that the maximal tumour dimension is relatively large (2.6 cm). However, the computed tumour volume is <1.5 cc due to the compression or ‘pancaking’ of the peripheral zone by the transition zone.

Table S1: AUC values for all variables at each endpoint.

Table S2: Prediction Accuracy of Selected Baseline Variables to Predict Future Outcomes.

Brisbane WG, Priester AM, Nguyen AV, Topoozian M, Mota S, Delfin MK, et al. Focal therapy of prostate cancer: Use of artificial intelligence to define tumour volume and predict treatment outcomes. BJUI Compass. 2025;6(1):e456. 10.1002/bco2.456