ADAPTING MOBILESAM FOR FEW-SHOT SEGMENTATION OF PROSTATE CANCER IN HISTOPATHOLOGY IMAGES

Abstract

Segmenting prostate cancer in tissue images is difficult because of irregular gland shapes, broken tissue structures, and very few labelled images available for training. This study introduces FrozenSE-SAM, a segmentation method that works well even with small datasets. It combines a frozen MobileSAM encoder with a lightweight decoder enhanced by Squeeze-and-Excitation (SE) blocks and is trained using Focal Tversky Loss, which helps focus on difficult regions. Unlike older methods that need extra shape information or lots of labels, FrozenSE-SAM can directly segment tumour regions without prompts. It was trained on only 35 tissue microarray (TMA) cores from the Gleason 2019 dataset and tested on 100 new samples. The model achieved a Dice score of 68.45%, which is better than U-Net (60.72%), Swin-UNETR (58.12%), and a Signed Distance Function (SDF) based model (62.77%). For measuring boundary accuracy, FrozenSE-SAM showed better performance with HD95 = 0.0228 mm and ASD = 0.0056 mm, compared to the SDF model (HD95 = 0.0328 mm, ASD = 0.0072 mm), and worse scores from U-Net and Swin-UNETR. Visual/Qualitative result also confirmed that FrozenSE-SAM was better at outlining complex tumour regions. It could accurately segment cribriform and fused glands without including nearby healthy tissue. In contrast, the SDF model produced blurry edges and missed finer structures, leading to under-segmentation. These results show that FrozenSE-SAM is a strong, reliable method for prostate cancer segmentation, especially in real-world situations with limited data.