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

Machine learning (ML), particularly deep learning (DL) and radiomics-based approaches, has emerged as a powerful tool for cancer outcome prediction using PET and SPECT imaging. However, the comparative performance of different techniques-handcrafted radiomics features (HRF), deep radiomics features (DRF), DL models, and hybrid fusion models (combinations of DRF, HRF, and clinical features)-remains inconsistent across clinical applications. This systematic review analyzed 226 studies published between 2020 and 2025 that applied ML to PET or SPECT imaging for cancer outcome prediction tasks. Each study was evaluated using a 59-item framework addressing dataset construction, feature extraction methods, validation strategies, interpretability, and risk of bias. We extracted key data, including model type, cancer site, imaging modality, and performance metrics such as accuracy and area under the curve (AUC). PET-based models (95%) generally outperformed SPECT, likely due to superior spatial resolution and sensitivity. DRF models achieved the highest mean accuracy (0.862 ± 0.051), while fusion models attained the highest AUC (0.861 ± 0.088). ANOVA revealed significant differences in accuracy (p = 0.0006) and AUC (p = 0.0027). Despite these promising findings, key limitations remain, including poor management of class imbalance (59%), missing data (29%), and low population diversity (19%). Only 48% adhered to IBSI (Image Biomarker Standardization Initiative) standards. This systematic review shows that DL and DRF-based models, especially in fusion with HRFs, outperform HRF-only methods for cancer outcome prediction using PET/SPECT, particularly in data-limited settings. Despite strong performance, challenges remain in interpretability and standardization, highlighting the need for unified DRF extraction frameworks across modalities.