New AI model LSDT improves prostate cancer MRI classification with 4-class system

A team of researchers led by Leyang Li has developed a new AI model called LSDT (Language-guided Segmentation-assisted Diagnostic Transformer) to improve prostate cancer diagnosis from multiparametric MRI. Current methods rely on PI-RADS assessment or simple binary classification (cancer vs. no cancer), which miss important pathological heterogeneity. To address this, they constructed the Prostate Cancer Histopathology Spectrum Dataset (PCa-HSD) that includes four clinically relevant classes: benign lesions, low-risk cancer, intermediate-risk, and high-risk disease. This approach better captures the spectrum of prostate pathology, especially benign lesions that are easily confused with cancer.

The LSDT model integrates two key innovations: zero-shot segmentation to provide anatomical priors (e.g., prostate gland boundaries) and effective multi-modal slice fusion for classification. By leveraging language-guided supervision from pathology reports, the model learns finer distinctions between risk levels. In five-fold cross-validation on a cohort of 344 patients, LSDT achieved a best average accuracy of 0.633 and a joint recall of 0.768, consistently outperforming baseline backbones. The model's performance demonstrates that combining pathological supervision with anatomical priors significantly enhances fine-grained classification.

This research represents a shift toward more clinically relevant AI diagnostics. Instead of binary cancer/no-cancer outputs, LSDT provides a risk stratification that mirrors actual clinical decision-making. The authors plan to release the code publicly, which could enable broader adoption in radiology workflows. The work is currently available on arXiv (2607.22703) and was submitted to a conference. If validated on larger cohorts, this approach could help radiologists prioritize high-risk patients and reduce unnecessary biopsies for benign lesions.