Voxel-wise supervised analysis of tumors with multimodal engineered features to highlight interpretable biological...

Area of Science:

• Radiomics and Medical Imaging Analysis
• Machine Learning in Oncology
• Translational Bioinformatics

Background:

• Deep learning models in medical imaging lack interpretability, hindering clinical translation.
• High-level feature abstraction in deep learning limits understanding of model predictions.
• Small datasets can cause instability in complex models like convolutional neural networks.

Purpose of the Study:

• To develop a radiomic model design that integrates interpretability of handcrafted features with sub-regional analysis.
• To enhance the clinical applicability of predictive and prognostic imaging models.

Main Methods:

• Utilized voxel-wise engineered radiomic features with average global aggregation and logistic regression.
• Applied the method to a small cohort (51 patients) of soft tissue sarcoma (STS) to predict lung metastasis.
• Employed positron emission tomography/computed tomography and MRI sequences for model building.

Main Results:

• Generated quantitative maps highlighting signal contributions within the tumor region of interest.
• Identified biological patterns in STS, including necrosis development and glucose metabolism, consistent with grading systems.
• Demonstrated the model's ability to spatially and quantitatively interpret radiomic decisions.

Conclusions:

• The proposed method enables spatial and quantitative interpretation of radiomic models.
• Facilitates sub-region identification and biological interpretation for improved patient stratification.
• Enhances the clinical utility of image-based predictive and prognostic models.