Supplementary Figures from Using Machine Learning Algorithms to Predict Immunotherapy Response in Patients with Advanced Melanoma
Year: 2023
Type: supplementary-materials
Abstract: Supplemental Figure 1. Training a Response Classifier to predict whether treatment-naïve advanced melanoma patients will respond to immunotherapy. (a) Schematic showing our computational workflow. We initially trained the neural network with a 5-fold cross validation. Hyper-parameter selection occurred during this phase. After selecting training conditions, we retrained the model on the entire NYU dataset using those parameters. We then tested the classifier on the independent cohort from Vanderbilt (b) During the 5-fold cross validation, neural network accuracy continued to improve until around 175,000 iterations. ROC curves show the performance of each of the 5 test sets at the optimal threshold. Supplemental Figure 2. Performance of the fully trained neural network and logistic regression classifiers from the NYU training dataset. Supplemental Figure 3. Development of a machine learning classifier that predicts whether treatment-naïve advanced melanoma patients will respond to immunotherapy or suffer disease progression. (a) Schematic showing the computational work-flow that is presented in this study. After development of the Segmentation Classifier, all tiles within the training and validation datasets were labeled for the 3 regions of interest. Hyper-parameter selection was performed during the 5-fold cross validation. Then, the model was retrained on the entire NYU dataset using the optimized foundation architecture and tested on the Vanderbilt cohort. (b) Suggested final pipeline where the slide is first pre-processed (tiled with suppression of the background tiles) and segmented by the first convolutional neural network. Identified tumor tiles are then assessed by the second network to extract a first estimate of the response to immunotherapy. The prediction is then refined by combining the output of the neural network with select clinical data. Supplemental Figure 4. Class activation mapping (CAM) analyses identify regions within the tiles that are used by the DCNN to classify the tile as (a) POD or (b) Response. From left to right, for each label, is an image of the tile image overlaid by the image from CAM analyses, the original tile image alone, and the image of CAM analyses alone. Supplemental Figure 5. CellProfiler was used to segment cell nuclei within tiles classified as having a high probability of POD or Response. The POD tiles analyzed had a prediction probability of 0.75 or higher (136,109 tiles). The Response tiles analyzed had a prediction probability of 0.25 or lower (51,220 tiles). Student's t-test was used to compare basic features of the segmented nuclei including area, density, and eccentricity. Supplemental Figure 6. Value importance of DCNN prediction, ECOG performance status, and treatment category was calculated as the absolute value of their Z-scores.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4361938323