# Multi-Task Deep Learning on MRI for Tumor Segmentation and Treatment Response Prediction in an Experimental Model of Hepatocellular Carcinoma

## TL;DR

A deep learning model was developed to segment liver tumors and predict treatment response using MRI scans in a rat model of liver cancer.

## Contribution

A multi-task deep learning framework was developed and biologically validated for tumor segmentation and treatment response prediction in hepatocellular carcinoma.

## Key findings

- The model achieved high tumor segmentation accuracy (Dice coefficient = 0.92) and treatment prediction (AUROC = 0.97).
- MRI-derived biomarkers strongly correlated with histological markers of tumor viability and apoptosis.

## Abstract

Background: Assessing the efficacy of combination therapies in hepatocellular carcinoma (HCC) requires both accurate tumor delineation and biologically validated prediction of therapeutic response. Conventional MRI-based criteria, which rely primarily on tumor size, often fail to capture treatment efficacy due to tumor heterogeneity and pseudo-progression. This study aimed to develop and biologically validate a multi-task deep learning model that simultaneously segments HCC tumors and predicts treatment outcomes using clinically relevant multi-parametric MRI in a preclinical rat model. Methods: Orthotopic HCC tumors were induced in rats assigned to Control, Sorafenib, NK cell immunotherapy, and combination treatment groups. Multi-parametric MRI (T1w, T2w, and contrast enhanced MRI) scans were performed weekly. We developed a U-Net++ architecture incorporating a pre-trained EfficientNet-B0 encoder, enabling simultaneous segmentation and classification tasks. Model performance was evaluated through Dice coefficients and area under the receiver operator characteristic curve (AUROC) scores, and histological validation (H&E for viability, TUNEL for apoptosis) assessed biological correlations using linear regression analysis. Results: The multi-task model achieved precise tumor segmentation (Dice coefficient = 0.92, intersection over union (IoU) = 0.86) and reliably predicted therapeutic outcomes (AUROC = 0.97, accuracy = 85.0%). MRI-derived deep learning biomarkers correlated strongly with histological markers of tumor viability and apoptosis (root mean squared error (RMSE): viability = 0.1069, apoptosis = 0.013), demonstrating that the model captures biologically relevant imaging features associated with treatment-induced histological changes. Conclusions: This multi-task deep learning framework, validated against histology, demonstrates the feasibility of leveraging widely available clinical MRI sequences for non-invasive monitoring of therapeutic response in HCC. By linking imaging features with underlying tumor biology, the model highlights a translational pathway toward more clinically applicable strategies for evaluating treatment efficacy.