Tandem: a Confidence-based Approach for Precise Medical Image Segmentation

Simone Monaco, Lorenzo Petrosino, Christodoulos Xinaris, Daniele Apiletti

Keywords: Medical image segmentation, small-objects detection, deep-learning models confidence

Abstract:

The advent of Healthcare 4.0 has heralded a disruptive change in medical diagnostics, with Artificial Intelligence (AI) playing a central role in improving diagnostic accuracy and treatment efficacy. This work addresses the integration of AI into medical imaging, specifically through the use of deep learning networks to analyze and segment medical images with high precision. Our study builds on the capabilities of U-Net-like models to address the challenges of segmenting clinical targets of different sizes, such as cysts and tumor masses, in medical images. We present a novel deep learning segmentation solution, Tandem (Tandem Analysis for Neural Detection and Evaluation Model), which combines a segmentation model with a classifier to produce a confidence map alongside the model's prediction.
This approach aims to improve segmentation accuracy by refining predictions and providing a mechanism to assess the reliability of the model, especially when identifying smaller clinically significant targets. We evaluate our method across various imaging modalities, including 2D and 3D acquisitions. We focus on detecting and segmenting kidney cysts associated with Autosomal Dominant Polycystic Kidney Disease (ADPKD) and tumor masses. The practical effectiveness of Tandem is demonstrated by the generation of reliable confidence maps that help clinicians make informed diagnostic and treatment decisions. This study represents a significant step towards precision medicine by improving the diagnostic capabilities of AI-driven systems in medical imaging.

Submission Number: 27