Interpretable Feed-Forward ML Model Achieves Dice >0.9 for Prostate Gland Segmentation on MRI

A novel feed-forward ML model (GUSL) matched or exceeded deep-learning performance for prostate gland segmentation on MRI, with a Dice score >0.9 across multiple datasets. Its interpretable, regression-based design avoids backpropagation and has a much smaller model size.

• Retrospective evaluation on two public and one private dataset; exact patient numbers not reported.
• GUSL uses a two-step pipeline for class imbalance and a boundary attention mechanism via residual correction, achieving state-of-the-art Dice among DL-based models.
• Interpretability claims lack clinical validation; small, potentially unrepresentative datasets and absence of external or prospective testing limit generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.