Data-efficient prostate cancer classification from T2-weighted MRI with mechanistic auditing and a dual-agent quality-control workflow

Diffusion-weighted images are sometimes unavailable or degraded in prostate MRI. We asked whether T2-weighted MRI alone retains usable cancer-related signal and whether the computation can be audited. An image-level transfer-learning ensemble reached out-of-fold AUC 0.883 (95% CI 0.836–0.924) on 243 de-identified institutional T2 images; a handcrafted image-level model reached 0.924 (0.889–0.953) and, in an exploratory external test of 137 Prostate158 patients, 0.643 (0.545–0.734) against 0.494 previously. A volume-level DINOv2-attention model reached AUC 0.717 (0.572–0.853) in 133 held-out biopsy-linked patients, 0.733 with test-time augmentation, and a locked precursor 0.618 (0.566–0.668) in 561 independent PROMIS patients. Distillation reproduced the frozen teacher with R ² 0.776, a transparent boosting machine 0.678 and a readable model tree 0.550, exposing dependence on gland size, low T2 signal and texture. A bounded dual-agent application gates input quality and anatomy before frozen inference, returning audit metadata. These findings support further evaluation of auditable T2-only analysis when complementary sequences are unavailable or degraded.