Baseline characteristics

In the training cohort (centers 1–3), 35.3% (84/238) and 64.7% (154/238) patients were diagnosed as PCa and benign (non-PCa), respectively and 11.3% (27/238) and 88.7% (211/238) as csPCa and non-csPCa, respectively. The positive/negative number and ratio concerning PCa and csPCa for each training and testing cohort are shown in Table 1. More details are in Supplementary Materials Section 6.

Diagnostic performance of DL models for predicting PCa and csPCa in PI-RADS 3 patients

To evaluate the performance of the proposed AttenNet models in predicting PCa and csPCa in PI-RADS 3 patients, an ablation experiment of the models with different modules was performed. As shown in Fig. 3, we compared the performance of models among the ResNet, ResNet combined with transfer module (ResNet-T), ResNet combined with transfer and channel attention modules (ResNet-TC), and AttentNet that referred to the ResNet combined with a transfer, channel attention, and soft attention modules. For the prediction of PCa, the AttenNet model achieved an area under the ROC curve (AUC) of 0.795 (95% CI: [0.700, 0.891]), 0.963 (95% CI: [0.915,1.00]), and 0.922 (95% CI: [0.810, 1.00]) in the external testing cohorts of center 4, center 5, and center 6, respectively (Fig. 3a). Furthermore, the AUC of the AttenNet model was significantly higher than those of the other models (i.e., ResNet, ResNet-T, and ResNet-TC) in center 4 (AUC = 0.661, 0.654, and 0.694, respectively) and center 5 (AUC = 0.640, 0.749, and 0.757, respectively) (Ps < 0.05, Fig. 3b). Although no difference in AUC was observed between the AttenNet (AUC = 0.922) and ResNet, ResNet-T, and ResNet-TC in center 6 (AUC = 0.611, 0.656, and 0.744, respectively) (Ps > 0.05, Fig. 3b), the AUC of the former was numerically higher than those of the latter. Thus, in terms of AUC, the AttenNet model showed the best performance for predicting PCa among these four DL models. Furthermore, in the external testing cohorts of center 4, center 5, and center 6, the AttenNet model achieved an ACC of 72.4% (71/98), 92.2% (59/64), and 82.6% (19/23); an SEN of 77.8% (28/36), 94.4% (17/18), and 100% (5/5); SPE of 69.4% (43/62), 91.3% (42/46), and 77.8% (14/18), respectively (Fig. 3a).

Fig. 3

Diagnosis performances of different deep learning models for predicting PCa and csPCa. a Diagnosis performances of the ResNet, ResNet-T, ResNet-TC, and AttenNet models for predicting PCa in three external testing cohorts. The AttenNet model yields the highest AUC compared with the other three models in predicting PCa. b ROC curves of the AttenNet model and the other three models for predicting PCa in three external testing cohorts. c Diagnosis performances of the ResNet, ResNet-T, ResNet-TC, and AttenNet models for predicting csPCa in two external testing cohorts. Similarly, the AttenNet model yields the highest AUC among all the models in predicting csPCa. d ROC curves of the AttenNet model and the other three models for predicting csPCa in two external testing cohorts. ROC, receiver operating characteristics; AUC, area under ROC curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; center 4, TZH, People’s Hospital of Taizhou; center 5, CSH, Changshu No.1 People’s Hospital; center 6, SKH, Suzhou Kowloon Hospital; PCa, prostate cancer; csPCa, clinically significant prostate cancer; ResNet-T, ResNet with transfer module; ResNet-TC, ResNet with transfer and channel attention modules; AttenNet, ResNet combined with transfer, channel attention and soft attention modules

As shown in Fig. 3c, for the prediction of csPCa, the AttenNet model achieved an AUC of 0.827 (95% CI: [0.703, 0.952]) and 0.926 (95% CI: [0.846, 1]) in the external testing cohorts of center 4 and center 5, respectively. As confirmed by the pathological exam, all PI-RADS 3 patients in the external testing cohort of center 6 were non-csPCa patients, and therefore the AUC for predicting csPCa in this cohort could not be calculated due to the lack of binary labels. Thus, the patients of this cohort were not used to test the prediction of csPCa. Further, in center 5, the AUC of the AttenNet model (AUC = 0.926) was significantly higher than those of the ResNet (AUC = 0.722, p = 0.037) and ResNet-T (AUC =0.775, p = 0.017), and marginally higher than that of ResNet-TC (AUC = 0.831, p = 0.063). Although no significant difference in AUC was observed between the AttenNet model (AUC = 0.827) and each of ResNet, ResNet-T, and ResNet-TC in center 4 (AUC = 0.669, 0.696, and 0.739, respectively) (Ps > 0.05, Fig. 3d), the AUC of the former was numerically higher than those of the latter (Fig. 3d). Thus, in terms of AUC, the AttenNet model showed the best performance for predicting csPCa among these four DL models. In addition, in the external testing cohorts of center 4 and center 5, the AttenNet model achieved an ACC of 73.5% (72/98) and 89.1% (57/64); a SEN of 86.7% (13/15) and 76.9% (10/13); a SPE of 71.1% (59/83) and 92.2% (47/51), respectively (Fig. 3c).

As revealed by DCA, the biopsy strategy based on the AttenNet model shows greater net benefit than that based on PI-RADS assessment (i.e., all PI-RADS 3 patients underwent the biopsy) for both detections of PCa (Fig. 4a) and csPCa (Fig. 4b) in the external testing cohorts. As shown in Fig. 4, the biopsy strategy based on the AttenNet model shows greater net benefit (Red line in Fig. 4) than that based on PI-RADS assessment (i.e., all patients with PI-RADS category 3 underwent the biopsy in the present study) (Blue line in Fig. 4) for detections of PCa (Fig. 4a) and csPCa (Fig. 4b) in each external cohort.

Fig. 4

Results of DCA of AttenNet for predicting PCa and csPCa. a Results of DCA of AttenNet for predicting PCa in three external testing cohorts. b Results of DCA of AttenNet for predicting csPCa in two external testing cohorts. The red lines indicate the net benefit of patients when using a biopsy based on the AttenNet model for detecting PCa (a) and csPCa (b). The blue lines indicate the net benefit of patients when they were all predicted to be positive (i.e., all patients with PI-RADS 3 underwent the biopsy) for the detection of PCa (a) and csPCa (b). The black lines indicate the net benefit of patients when they were all predicted to be negative for the detection of PCa (a) and csPCa (b)

Clinical practice of the AttenNet models for predicting PCa and csPCa in PI-RADS category 3 patients

Figure 5 shows more details of the prediction results of AttenNet models from the clinical practice perspective. As shown in Fig. 5a, in each external testing cohort, the PI-RADS 3 patients were upgraded to PI-RADS 3U and downgraded to PI-RADS 3D according to the prediction results of the PCa by AttenNet model (i.e., PCa or non-PCa), respectively. The detailed results in each external testing cohort are described in Supplementary Section 6. As shown by Fig. 5a, 69.4% (Fig. 5a, center 4: 43/62) to 91.3% (Fig. 5a, center 5: 42/46) of benign (i.e., non-PCa) patients were identified by our AttenNet model from PI-RADS 3 patients of the external testing cohorts. In other words, these benign patients would have been spared from various clinical therapies and anxieties if the AttenNet model had been used to diagnose PCa.

Fig. 5

Clinical practice of AttenNet models for predicting PCa and csPCa in PI-RADS category 3 patients. a The downgrading and upgrading results of PI-RADS category 3 patients using the AttenNet model for predicting PCa in three external testing cohorts of center 4, center 5, and center 6. b The downgrading and upgrading results of PI-RADS category 3 patients using the AttenNet model for predicting csPCa in two external testing cohorts of center 4 and center 5. PI-RADS, prostate imaging-reporting and data system version; PI-RADS 3U, PI-RADS category 3 upgrade; PI-RADS 3D, PI-RADS category 3 downgrade; center4, CSH, Changshu No.1 People’s Hospital; center5, TZH, People’s Hospital of Taizhou; center6, SKH, Suzhou Kowloon Hospital; PCa, prostate cancer; ciPCa, clinically insignificant prostate cancer; csPCa, clinically significant prostate cancer

As shown in Fig. 5b, in each external testing cohort, PI-RADS 3 patients were upgraded and downgraded to PI-RADS 3U and PI-RADS 3D according to the prediction results of csPCa by AttenNet model (i.e., csPCa and non-csPCa), respectively. The detailed results in each external testing cohort were described in Supplementary Section 7. As shown in Fig. 5b, 71.1% (Fig. 4b, center 4: [48 + 11]/83) to 92.2% (Fig. 5b, center 4: [43 + 4]/51) of non-csPCa were identified by our AttenNet model from the equivocal PI-RADS 3 patients of the external testing cohorts, all of whom had undergone biopsies because these non-csPCa patients could not be identified using PI-RADS assessment. Those results suggest that our AttenNet model has the potential to reduce unnecessary biopsies for non-csPCa patients.

The results of subgroup analysis with AttenNet models for different levels of tumor size and PSA

Additionally, we assessed the performance of the AttenNet models for predicting PCa and csPCa in subgroups of PI-RADS 3 patients according to different levels of tumor size and PSA. As revealed by the subgroup analysis (Fig. 6), the AttenNet models for predicting PCa and csPCa achieved satisfactory performance in different levels of D-max and PSA (except for the subgroup of 0 ≤ PSA < 10 ng/mL for predicting csPCa in the center 4) (Supplementary Section 8).

Fig. 6

The results of subgroup analysis with AttenNet models for different levels of tumor size and PSA. a The performance of the AttenNet model for predicting PCa in each subgroup. b The performance of the AttenNet model for predicting csPCa in each subgroup. Center 4, CSH, Changshu No.1 People’s Hospital; center 5, TZH, People’s Hospital of Taizhou; center 6, SKH, Suzhou Kowloon Hospital; PCa, prostate cancer; csPCa, clinically significant prostate cancer; PSA, prostate-specific antigen; ROC, receiver operating characteristics; AUC, area under ROC curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; D-max, diameter in greatest dimension