Impact and examples

The impact of machine learning on medicine and healthcare continues to increase. Many examples in literature or in the news focus on singular models making new progress on specific problems. However, the task of providing a service to a large number of potential patients requires a broad based solution. In particular, to enter production it will need to be one that is robust against differences in patient background or available information while still offering the competitive performance.

This project provides a minimum viable product for a best performing, end-to-end pipeline that is able to take a variety of patient information and provide a prostate cancer diagnosis. With over 1100 patients in the project dataset featuring a variety of healthcare institutions, medical procedures, and information available about each individual, this product is much more likely to be robust to new clinical settings and patients. In addition, the use of uncertainty estimation can provide context for when a patient’s status may be ambiguous and flagged for a manual follow up.

Specifically, MRI Analyzer compares to recent publications in the following way:

Note that the AUC-ROC reported above is found from the full curve, as used in other literature examples. Elsewhere on this site both the full curve and single point value from the 50% decision boundary are reported.

use case #1: flexible pipeline

MRI Analyzer’s larger, more diverse dataset, coupled with robust training, makes it better prepared to treat patients from a wide variety of backgrounds and in differing circumstances. As described on the Dataset page, various scanning procedures and equipment were used and differed from patient to patient. This variety both makes the pipeline more robust for patients in across backgrounds in the future, but also likely helped to prevent overfitting to the training set currently. 

Further, a patient only needs an MRI and Ultrasound scan in order for a classification to be made in line with model training and evaluation. All other variables, including levels of Prostate Specific Antigen (PSA) from blood tests and all patient metadata, were imputed with the training group’s average whenever missing for a patient. Therefore, the best performing AUC-ROC metric includes these cases of partial missing data. While only future evaluation on new patients will reveal whether this performance continues, receiving competitive scores on the current dataset with this method indicates that this flexibility is within reach for patients.


use case #2: uncertainty quantifier

The inclusion of full-pipeline uncertainty estimates not only provides better context for an individual cancer prediction, but also can be used to improve processing of patients in bulk.

The image above demonstrates how an individual with an ambiguous point estimate can be better understood with an accompanying uncertainty estimate. In this case, one patient had a predicted probability just under ~40%. For the simple scenario of binary classification, a naive baseline may be to flag anyone between 30-70% because of their proximity to the decision boundary. In this case, however, this individual was correctly classified despite their near-ambiguous point estimate. This is reflected with a quantified uncertainty estimate that is much lower than its neighbors in this same baseline, which were incorrectly classified. Although there are visible examples of patients which were both certain and incorrectly classified, uncertainty can still demonstrate better filtering of ambiguous patients.


use case #3: Prioritization tool


Many elective operations and surgeries were postponed or cancelled during the COVID-19 pandemic. This meant that all of the patients whose doctors were suspecting them of having cancer had to postpone their biopsy operations. As a result, patients who were having symptoms of prostate cancer had to wait until COVID-19 had passed in their area or until their symptoms worsened. To better serve patients who are most in need, doctors can use MRI Analyzer as a complementary tool to predict point estimate probabilities of cancer along with uncertainty estimates on a patient by patient basis.

For example, a patient goes to the doctor since he is making very frequent bathroom visits and has seen blood in their urine. The doctor takes blood samples and also sends the patient for an MRI scan. Due to COVID-19, many patients cannot proceed beyond the MRI scan without further certainty of their condition. Instead of waiting for worsening symptoms, MRI Analyzer can be a supplementary tool to determine the cancer status along with a quantification of its uncertainty for the patient. In particular, for scenarios with high demand for medical resources such as COVID-19, MRI Analyzer makes the evaluation process more efficient so doctors can spend time better focused on patients and situational needs.


[1] Sonn GA, Fan RE, Ghanouni P, Wang NN, Brooks JD, Loening AM, Daniel BL, To’o KJ, Thong AE, Leppert JT. Prostate Magnetic Resonance Imaging Interpretation Varies Substantially Across Radiologists. Eur Urol Focus. 2019 Jul;5(4):592-599. doi: 10.1016/j.euf.2017.11.010. Epub 2017 Dec 7. PMID: 29226826.


[2] Minh Hung Le et al 2017 Phys. Med. Biol. 62 6497

[3] Yoo, S., Gujrathi, I., Haider, M.A. et al. Prostate Cancer Detection using Deep Convolutional Neural Networks. Sci Rep 9, 19518 (2019).