Lesion Segmentation and Classification on Prostate MRI Using Deep Learning and Haralick Texture Features

We present a deep learning (DL) framework integrating Haralick texture features (HTF) to enhance lesion detection and characterization, crucial for prostate cancer diagnosis and treatment planning. A custom ProstateX dataset of 3D T2-weighted MR images (T2WI) from 344 patients was resampled and used in the study. A modified Retina U-Net served as the backbone for class-independent lesion segmentation, followed by bounding box regression to refine detection and classification for lesion categorization. Axial T2WI were augmented with HTF maps (Contrast, Energy, Correlation, Homogeneity) computed from 2D slices and incorporated as multi-channel inputs for segmentation and aggressiveness prediction using Gleason scores as ground truth. On the ProstateX test set, our model achieved strong performance (AUC = 0.87, accuracy = 0.806, sensitivity = 0.923, specificity = 0.776) and generalized well on additional multi-center datasets. These results highlight the promise of combining DL with HTF to improve prostate cancer lesion detection and classification.