Lesion detection using artificial intelligence models in MR images of prostate cancer and prostatitis patients and comparison of model performance.

Aim: The diagnosis of prostate cancer and prostatitis becomes challenging when using biparametric Magnetic Resonance (MR) images. This research investigates deep learning models to assess their capability for improving diagnostic accuracy and assisting radiologists. Methods: This retrospective study analyzed 153 patients who received histopathological diagnoses of prostate cancer or prostatitis between January 2017 and December 2023. Patients were categorized according to PI-RADS scores, and both T2A and ADC-DWI (Apparent Diffusion Coefficient-Diffusion-Weighted Imaging) sequences were examined. Expert radiologists labeled the images prior to lesion detection with the Faster R-CNN (Faster Region-based Convolutional Neural Network) model. Nine different classification models were trained using normal and augmented datasets to evaluate their performance. Model reliability was further assessed through cross-validation and statistical significance testing. Results: The Faster R-CNN model achieved 96% accuracy (95% CI: 93.2-98.8%) for P5 and 99% accuracy (95% CI: 96.7-100%) for prostatitis in T2A sequences, and 90% accuracy (95% CI: 85.4-94.6%) for P5 and 97% accuracy (95% CI: 93.8-100%) for prostatitis in ADC-DWI sequences. However, the model failed to effectively detect P4 lesions (0% sensitivity in T2A and 30% in ADC-DWI). The model demonstrated comparable performance to expert radiologists, with no significant difference in overall P5 detection (p > 0.05), and Cohen's kappa indicated substantial agreement (κ = 0.86). The classification models achieved up to 97% accuracy with InceptionV3 in T2A sequences and up to 99% accuracy with DenseNet201 in ADC-DWI sequences. To further evaluate discriminative performance, AUROC values were calculated for all classification models. In T2A sequences, AUROC scores were DenseNet201 (0.98), EfficientNetV2L (0.99), InceptionV3 (0.99), MobileNetV2 (0.92), NASNetLarge (0.83), ResNet50 (0.76), VGG16 (0.98), VGG19 (0.97), and Xception (0.96). In ADC-DWI sequences, AUROC values were DenseNet201 (0.99), EfficientNetV2L (0.96), InceptionV3 (0.99), MobileNetV2 (0.82), NASNetLarge (0.90), ResNet50 (0.64), VGG16 (0.96), VGG19 (0.86), and Xception (0.97), reinforcing the superior discriminative ability of DenseNet201 and InceptionV3 across modalities. Conclusion: The deep learning models demonstrated promising diagnostic capabilities, comparable to radiologists, in distinguishing prostatitis and P5 prostate cancer lesions. Overall, the findings suggest that AI-based diagnostic tools hold potential as clinical decision support systems.