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

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Histopathology provides cellular diagnosis but is invasive and offers limited tumor scale.
• Medical imaging (X-ray, CT, MRI, PET) offers whole-tumor characterization but traditionally relied on subjective human interpretation.
• Radiomics, applying machine learning (ML) to medical images, aims to uncover hidden disease patterns.

Purpose of the Study:

• Critically discuss the trajectory of ML in radiomics for oncology.
• Identify systemic weaknesses hindering clinical translation of ML in radiomics.
• Propose future directions for developing robust ML systems in oncology.

Main Methods:

• Review and analysis of studies employing various imaging modalities in oncology ML.
• Exploration of innovative ML algorithms and their achievements.
• Identification of challenges impacting model generalizability and interpretability.

Main Results:

• Two major challenges identified: inter-modality/inter-scanner variability affecting model generalizability, and 'interpretability gaps' in ML decision-making.
• These challenges impede the clinical translation of even advanced ML algorithms.
• Current ML models often perform well in labs but struggle in real-world clinical settings.

Conclusions:

• Variability and lack of interpretability are significant barriers to clinical adoption of ML in radiomics.
• Future ML development should prioritize performance in real-world settings over lab-based achievements.
• Addressing these systemic weaknesses is crucial for advancing precision oncology through AI-driven imaging analysis.