Machine Learning for Radiomics in Oncology: Challenges, Limitations, and Future Directions

Area of Science:

• Oncology
• Radiology
• Machine Learning
• Medical Imaging

Background:

• Precision oncology integrates histopathology and medical imaging for tumor characterization.
• Histopathology provides cellular diagnosis but is invasive and offers limited tumor scale.
• Medical imaging (X-ray, CT, MRI, PET) offers comprehensive tumor visualization but traditionally relies on subjective human interpretation.

Purpose of the Study:

• To critically discuss the trajectory of machine learning (ML) in radiomics for oncology.
• To identify systemic weaknesses hindering the clinical translation of ML in radiomics.
• To explore studies across various imaging modalities to assess ML applications in oncology.

Main Methods:

• Review and analysis of existing studies applying ML to radiomics in oncology.
• Exploration of different medical imaging modalities (e.g., X-ray, CT, MRI, PET).
• Critical evaluation of ML algorithm performance and generalizability in real-world settings.

Main Results:

• Identified two major challenges: inter-modality/inter-scanner variability affecting model generalizability.
• Highlighted 'interpretability gaps' in understanding ML decision-making processes.
• Observed that these challenges impede the clinical translation of even advanced ML algorithms.

Conclusions:

• Significant variability in imaging data and lack of ML model interpretability are critical barriers.
• Future ML development in radiomics must prioritize performance in real-world clinical settings.
• Advocating for robust, generalizable, and interpretable ML systems for enhanced oncology applications.