Handcrafted vs. Deep Radiomics vs. Fusion vs. Deep Learning: A Comprehensive Review of Machine Learning -Based Cancer...

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Oncology
• Radiomics and Deep Learning

Background:

• Machine learning (ML), including deep learning (DL) and radiomics, is increasingly used for cancer outcome prediction with PET and SPECT imaging.
• Inconsistent performance of different ML techniques (handcrafted radiomics features (HRF), deep radiomics features (DRF), DL, and hybrid fusion models) necessitates a comparative analysis.

Purpose of the Study:

• To systematically review and compare the performance of various ML techniques for cancer outcome prediction using PET/SPECT imaging.
• To identify the most effective ML approaches and highlight existing limitations and challenges in the field.

Main Methods:

• Systematic review of 226 studies (2020-2025) applying ML to PET/SPECT for cancer outcome prediction.
• Evaluation using a 59-item framework covering dataset construction, feature extraction, validation, interpretability, and bias.
• Data extraction included model type, cancer site, imaging modality, accuracy, and AUC.

Main Results:

• PET-based models generally outperformed SPECT.
• Deep radiomics features (DRF) models achieved the highest mean accuracy (0.862 ± 0.051).
• Fusion models (DRF, HRF, clinical data) attained the highest AUC (0.861 ± 0.088), with significant differences observed in accuracy and AUC (p < 0.003).

Conclusions:

• Deep learning and DRF-based models, particularly when fused with HRFs, outperform HRF-only methods for cancer outcome prediction with PET/SPECT.
• Significant limitations include poor class imbalance management, missing data, low population diversity, and lack of adherence to standardization initiatives like IBSI.
• Further research is needed to address interpretability and standardization, advocating for unified DRF extraction frameworks.