A Hybrid Deep Learning Approach for Liver Tumor Segmentation Using DeepLabV3+ and Hidden Markov Models

The liver, being the largest solid organ in the human body, is one of the most important players in metabolism and digestion processes. It is also a site where both primary and secondary tumors — originating from distant organs such as the lungs or other abdominal parts such as the pancreas and colon — can originate or metastasize. Therefore, the liver is regularly screened for the presence of lesions. These lesions require precise segmentation techniques to accurately diagnose cancer and improve patient monitoring, disease progression, and response to treatment. In this paper, a slightly modified version of a DeepLabV3+ network, a well–known and state–of–the–art segmentation model, paired with a Hidden Markov Model (HMM) based noise reduction module, is employed and trained on the Medical Segmentation Decathlon (MSD) liver tumor data set. This collection of liver lesions is a fraction of the MSD international challenge dedicated to identifying a general– purpose algorithm for medical image segmentation. The model is then evaluated on the test set of the same dataset with pixel–pixel accuracy and Intersection over Union (IoU).