Prostate cancer (PCa) is one of the most commonly diagnosed cancers in elderly men and the sixth leading cause of cancer-related death in Japan. Prostate-specific antigen (PSA) testing is widely used for early detection and has helped reduce cancer-related deaths, but it has a critical blind spot.
Men with low to intermediate PSA levels - below 20 ng/mL - can still harbor aggressive, clinically significant cancers. PSA alone has poor discrimination in this range, meaning that many patients with life-threatening tumors could be missed or undertreated, while others with insignificant tumors may receive unnecessary biopsies.
Clinically significant prostate cancer - defined as ISUP grade group 2 or higher - requires definitive treatment and can be life-threatening if not identified early. Treatment options include radical prostatectomy, radiation therapy, brachytherapy, and hormonal therapy, many of which are most effective when cancer is caught before it progresses.
This study aimed to improve prediction of clinically significant PCa specifically in the low to intermediate PSA range by combining multiple data types - ultrasound imaging, MRI, and clinical laboratory values - through a multimodal deep learning approach, predicting cancer risk before biopsy is performed.
The study enrolled 178 consecutive patients who underwent ultrasound-guided prostate biopsy at Nippon Medical School Hospital in Tokyo between August 2019 and June 2020. After exclusions, 151 patients were included in the analysis, with a median age of 71 years.
Each patient underwent biparametric MRI before biopsy, with scans performed on both 1.5 and 3.0 tesla scanners. Two key MRI sequences were collected: T2-weighted imaging (T2WI), which shows anatomical detail of the prostate, and diffusion-weighted imaging (DWI) with its derived apparent diffusion coefficient (ADC) maps, which reflect tissue cellularity.
Ultrasound imaging was performed at four prostate locations (base, middle, middle-apex, and apex) using a transrectal probe. In total, the analysis covered 583 ultrasound images, 1,540 T2WI images, and 1,487 DWI or ADC images. All images were standardized to 256x256 pixels for deep learning input.
Histopathological evaluation was conducted independently by two pathologists using the ISUP grading system, with clinically significant PCa defined as ISUP grade groups 2 through 5. Of the 151 included cases, 57% were diagnosed with clinically significant PCa.
The AI pipeline used a deep convolutional neural network pre-trained on ImageNet (specifically the Xception architecture) to process each imaging modality independently. Rather than using all images from a patient, the system automatically selected the three most informative images per modality based on the highest predicted probability scores.
For each modality, the three selected images generated predicted probability scores from the neural network. These probabilities, representing the likelihood of clinically significant PCa in each image, were then combined as feature inputs to a support vector machine (SVM) classifier for the final integrated prediction.
The integrated step took 12 probability values (three images from each of four imaging modalities) plus the clinical PSA value, feeding all 13 features into the SVM. This two-step architecture - deep learning for feature extraction, SVM for final classification - allowed the model to leverage both image complexity and structured clinical data.
Data augmentation techniques including zoom range adjustments were applied to increase training data diversity. The dataset was split chronologically: cases from August 2019 to February 2020 formed the training set (112 cases) and cases from March to June 2020 formed the test set (39 cases).
When evaluated across all 151 patients with no PSA restriction, individual modalities showed modest predictive performance: PSA achieved an AUC of 0.649, ultrasound imaging 0.715, T2WI 0.738, diffusion-weighted imaging 0.582, and ADC maps 0.690. However, the integrated multimodal analysis achieved an AUC of 0.878 (95% CI: 0.772-0.984), significantly surpassing PSA alone (p = 0.024).
The critical test was the PSA-restricted analysis - when only patients with PSA below 20 ng/mL were evaluated (122 patients). In this group, PSA's discriminatory ability dropped to an AUC of only 0.574. T2WI performed best among individual modalities at 0.803, but the integrated model still achieved the highest AUC of 0.862 (95% CI: 0.723-1.000, p = 0.032 vs. PSA alone).
These results confirm the core hypothesis: even when PSA provides limited information, combining multiple imaging modalities through deep learning can maintain high predictive accuracy for clinically significant prostate cancer.
Both false-negative and false-positive cases in the PSA-restricted group involved patients with PSA levels below 10 ng/mL and MRI features consistent with benign prostatic hyperplasia. This highlights that hyperplasia patterns - which can mimic cancer features - remain a key diagnostic challenge for AI models in this PSA range.
This study is distinctive in focusing specifically on patients with low to intermediate PSA levels, a group where existing biomarkers and single-modality imaging tools often fail. Previous multimodal AI studies using multiparametric MRI showed promise, but did not specifically target this challenging PSA range.
The approach compares favorably to emerging biomarker tools. The Prostate Health Index (PHI), a commercially available blood test that combines multiple PSA variants, achieves a combined sensitivity of 0.874 with specificity of 0.569 for detecting clinically significant PCa. The multimodal AI here achieved comparable sensitivity while requiring no additional blood tests beyond standard PSA.
A key design principle was clinical practicality: the system only uses medical data already collected in routine prostate cancer workups - PSA values, ultrasound, T2WI, DWI, and ADC. This means it could be integrated into existing clinical workflows without requiring additional procedures or tests, reducing adoption barriers.
The benign prostatic hyperplasia pattern emerged as a key confounder, appearing in both false-negative and false-positive cases. Adding more training images that capture hyperplasia-cancer distinctions is a clear next step for improving model accuracy in this boundary region.
The clinical goal of this work is to help urologists decide who needs a prostate biopsy before the procedure is performed. By predicting clinically significant cancer risk from pre-biopsy data, the AI could help reserve biopsies for those most likely to benefit while avoiding unnecessary invasive procedures in low-risk men.
Treatment planning for prostate cancer - whether radical prostatectomy, radiation therapy, brachytherapy, or hormonal therapy - depends critically on accurate risk stratification. Identifying clinically significant cancer early, before it progresses, is when these treatments are most effective and when the widest range of options remains available.
The authors emphasize that AI should support, not replace, physician judgment. The system is designed to work within existing medical workflows, providing additional predictive information that urologists can use alongside their clinical assessment rather than as a standalone diagnostic tool.
Integration of AI into prostate cancer management pathways could help personalize decisions: men identified as high-risk for clinically significant cancer could be fast-tracked to biopsy, while those at low risk might be monitored with less invasive follow-up, reducing unnecessary procedures and healthcare costs.
This study demonstrates that multimodal deep learning can significantly improve prediction of clinically significant prostate cancer in patients with low to intermediate PSA levels - precisely the group where standard PSA testing performs worst.
By combining ultrasound, T2WI, DWI, ADC, and PSA values through a two-stage AI pipeline, the system achieved an AUC of 0.862 in the PSA-restricted group, compared to PSA's AUC of just 0.574 in the same group. The improvement is substantial and clinically meaningful.
The main limitation is sample size - 151 patients from a single institution. External validation at larger, multi-site cohorts is needed before wider clinical deployment. Future work will also include transperineal biopsy cases and aim to improve handling of benign prostatic hyperplasia patterns.
As AI tools become more integrated into oncology workflows, approaches like this one that leverage existing routine clinical data across multiple modalities offer a practical path to better-personalized cancer management - without requiring new tests, new equipment, or fundamental changes to how patients are cared for.