Table 1.
Prostate cancer review articles summary
| Legend | Name/Type of Model | No. of patients | Cancer | Summary | Data Analysis | Implications | References |
|---|---|---|---|---|---|---|---|
| 1-A | Random forest | 7543 | Prostate Cancer | A clinical registry of relevant prostate cancer data across 45 urology practices in the state of Michigan, a random forest machine learning (ML) model was developed to predict probability of receiving a given treatment based on clinical, pathologic, and demographic factors. | Accuracy AUC was 0.81 | This web- and smartphone-based platform can serve as a useful tool for patients to better understand treatment options and decisions and be able to take charge of their care. | 18. Auffenberg et al. |
| 1-B | CNN | 886 (641 for training, 245 for testing) | Prostate Cancer | A CNN model was developed to automatically assign Gleason scores to H&E stained prostate cancer tissue microarrays. | Inter-annotator agreement between the model and 2 pathologists using Cohen’s quadratic kappa statistic was kappa = 0.75 and 0.71 respectively. | Deep learning technology has the potential to assist pathologists in histopathologic grading of prostate cancer and mitigate the effects of inter-pathologist variability. | 19. Arvaniti, E., et al. |
| 1-C | CNN | 38 patients (96 biopsies) | Prostate Cancer | A CNN model was developed to detect Gleason patterns (GP) and determine grade groups (GG) from H&E stained tissue sections of 96 prostate biopsies. | The model’s ability to differentiate between malignant (GP ≥ 3) and non-malignant (GP < 3) sections reached an accuracy of 92% (F-score = 0.93). Automated GG reached a concordance of 65% with a genitourinary pathologist (kappa = 0.70). | Computer-aided automated histopathologic grading of prostate cancer has the potential to reduce both inter-pathologist variability and time required to diagnose. | 20. Lucas, M. et al. |
| 1-D | SVM | 147 | Prostate Cancer | Three ML algorithms were tested to be able to classify prostate cancer Gleason scores based on combined apparent diffusion coefficient (ADC) and T2-weighted prostate MRIs of patients with biopsy-proven prostate cancer undergoing radical prostatectomy. ML results were validated with Gleason grade patterns obtained from histopathological analysis of the excised prostates. | Distinguishing between GS 6 vs. GS ≥ 7 resulted in 93% accuracy for cancers occurring in both peripheral (PZ) and transition (TZ) zones. Distinguishing between GS 7 (3+4) vs. GS 7 (4+3) resulted in 92% accuracy for cancers occurring in both PZ and TZ. | The ability of ML algorithms to differentiate Gleason scores from multiparametric MRI images presents a unique method for non-invasive and accurate diagnosis of prostate cancer, potentially decreasing the need for prostate biopsies. | 22. Fehr, D. et al. |
| 1-E | k-nearest neighbor | 150 | Prostate Cancer | ML algorithm was developed to generate low-dose-rate brachytherapy treatment plans for patients with prostate cancer. Automatically generated plans were compared to both expert radiation therapist (RT) and brachytherapist (BT) plans. | Dosimetry and clinical quality of ML plans were found to be equivalent to RT and BT plans. Planning time was significantly reduced with the ML plan (mean 0.84 min vs. 17.88 min for RT/BT, p = 0.020). | The capability of the ML algorithm to reach equivalent clinical quality to experts in brachytherapy planning combined with its significant reduction in time and resource expenditure is a promising utilization of ML for aiding in planning treatments. | 23. Nicolae, A. et al. |
| 1-F | CNN | 418 for training | Prostate Cancer | An automatic contour propagation pipeline was created using deformable image registration (DIR) to develop online adaptive intensity-modulated proton therapy (IMPT) plans for prostate cancer. | A conservative success rate of 80% was achieved, signifying that 80% of plans generated could be used without manual correction. | IMPT is capable of delivering a localized dose of radiation to target tissue, while minimizing damage to surrounding organs. However, IMPT is more sensitive to daily anatomical changes, exacerbating treatment-associated toxicity. DIR and deep learning have the capability to create an automated plans that adapt to these daily anatomical changes and protect organs at risk (OARs). | 24. Elmahdy, M.S., et al. |
| 32 for testing |