Stephen X. Skapek

University of Texas Southwestern Medical Center
Country flagUnited States

Research Interests

ApplyKite TwinLearns as you apply

Make your next move.

Prepare your application

Review drafts before you send them.

Explore related searches

Articles (10)

Prediction of RAS pathway mutations using a trained CNN. A, Workflow for deep learning of RAS pathway mutations from FN-RMS WSIs. B and C, Representative ( B ) H&E images and ( C ) class activation maps of a RAS pathway wild-type tumor and a tumor with a KRAS p.G12C mutation (VAF = 0.659). D, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. E, Statistics for confusion matrix. F, Average ROC curve from holdout test data.

Year:

2025

Supplemental Figure S3. Sample partitioning for training and testing a RAS pathway mutation predictive model using K-fold cross-validation. Three independent experiments were trained on a random selection of samples for training, validation and testing.

Year:

2025

Deep learning of histologic features from RMS tumor tissue. A, (i) Samples are randomly selected for training, validation, or holdout test groups with k-fold cross validation. Networks were trained to recognize (ii) basic histological characteristics and (iii) features associated with FOXO1 fusion status, or (iv) other relevant RMS mutations. (v) A predictive model was also developed to predict disease risk using only an H&E image. B, Representative H&E images (left), expert pathologist manual annotation (middle), and pixel-level segmentation results of the A.I. algorithm (right). C, Histogram of the weighted IoU scores from holdout test data ( n = 29). Samples corresponding to B are indicated. D, Average and weighted intersection over union (IoU) scores of A.I. performance across a 3-fold cross-validation set compared with a pathologist manual annotation.

Year:

2025

Supplemental Table S1. Whole slide image tissue segmentation statistics by an expert pathologist and probability prediction using a trained convolutional neural network. Supplemental Table S2. Clinical and molecular characteristics of FN-RMS samples used for training models for mutation prediction. Yellow boxes indicate genes included in defining the RAS pathway. Supplemental Table S3. Clinical information with COG risk stratification of FN-RMS samples used for training a prognostication predictive CNN.

Year:

2025

TP53 gene mutation prediction from H&E images. A, Workflow for deep learning of TP53 mutations from FN-RMS WSIs. B and C, Representative ( B ) H&E images and ( C ) class activation maps of a TP53 wild-type tumor and a tumor with a TP53 p.P278T mutation (VAF = 0.474). D, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. E, Performance statistics for TP53 mutation prediction. F, Average ROC curve for TP53 mutation prediction using holdout test data. G, Dot plot of TP53 mutant samples ( n = 15) showing relationship between TP53 mutation VAF and A.I. positive prediction probability. Statistical analysis was performed using the Mann–Whitney U test.

Year:

2025

Fusion status prediction from H&E images. A, Workflow for deep learning of normal tissue, FN-RMS, and FP-RMS. B and C, Representative ( B ) H&E images and ( C ) class activation maps as predicted by the trained model. D, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. E, Performance statistics for FP-RMS and FN-RMS prediction. Note: Normal tissues ( n = 15) were classified with 100% sensitivity and 100% specificity and are included in performance metric calculations. F, Average ROC curve for FP-RMS vs. FN-RMS vs. normal predictions using holdout test data. G, Workflow for testing the trained algorithm against an independent RMS TMA dataset. H, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. I, Performance statistics for FP-RMS and FN-RMS predictions on an independent dataset. J, Average ROC curve for FP-RMS vs. FN-RMS using an independent RMS tissue dataset.

Year:

2025

MYOD1 gene mutation prediction from H&E images. A, Workflow for deep learning of MYOD1 mutations from FN-RMS WSIs. B and C, Representative ( B ) H&E images and ( C ) class activation maps of a MYOD1 wild-type tumor and a tumor with a MYOD1 p.L122R mutation (VAF = 0.919). D, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. E, Performance statistics for MYOD1 mutation prediction. F, Average ROC curve for MYOD1 mutation prediction using validation data. G, Plot of MYOD1 mutation positive ratio thresholds with test datasets (blue = known MYOD1 wild-type; red = known MYOD1 mutant) and independent dataset of known MYOD1 -mutant tumors n = 10 (green) shown as geometric mean ± 1 geometric SD.

Year:

2025

Deep learning algorithms can predict event-free and overall survival from H&E histology alone. A, Schematic for training a convolutional neural network for survival. Patients with FN-RMS ( n = 264) were partitioned into three groups based on EFS length to ensure even sampling and fed into a deep learning architecture with a Cox proportional hazard layer to assess risk. B, Kaplan–Meier curves for EFS based on COG clinical risk group assessment. C, Kaplan–Meier curves for event-free survival based on A.I. hazard prediction. D, Kaplan–Meier curves for OS based on COG clinical risk group assessment. E, Kaplan–Meier curves for OS based on the same A.I.-predicted risk grouping as in C .

Year:

2025

Purpose: Rhabdomyosarcoma (RMS) is an aggressive soft-tissue sarcoma, which primarily occurs in children and young adults. We previously reported specific genomic alterations in RMS, which strongly correlated with survival; however, predicting these mutations or high-risk disease at diagnosis remains a significant challenge. In this study, we utilized convolutional neural networks (CNN) to learn histologic features associated with driver mutations and outcome using hematoxylin and eosin (H&E) images of RMS. Experimental Design: Digital whole slide H&E images were collected from clinically annotated diagnostic tumor samples from 321 patients with RMS enrolled in Children's Oncology Group (COG) trials (1998–2017). Patches were extracted and fed into deep learning CNNs to learn features associated with mutations and relative event-free survival risk. The performance of the trained models was evaluated against independent test sample data ( n = 136) or holdout test data. Results: The trained CNN could accurately classify alveolar RMS, a high-risk subtype associated with PAX3/7-FOXO1 fusion genes, with an ROC of 0.85 on an independent test dataset. CNN models trained on mutationally-annotated samples identified tumors with RAS pathway with a ROC of 0.67, and high-risk mutations in MYOD1 or TP53 with a ROC of 0.97 and 0.63, respectively. Remarkably, CNN models were superior in predicting event-free and overall survival compared with current molecular-clinical risk stratification. Conclusions: This study demonstrates that high-risk features, including those associated with certain mutations, can be readily identified at diagnosis using deep learning. CNNs are a powerful tool for diagnostic and prognostic prediction of rhabdomyosarcoma, which will be tested in prospective COG clinical trials.

Year:

2025

Year:

2023

Collaborators (6)

Jun Wei

National Cancer Institute

UNITED STATES
View details

Corinne M. Linardic

Duke University

UNITED STATES
View details

Douglas Hawkins

Professor

University of Washington

UNITED STATES
View details

Javed Khan

National Institutes of Health

UNITED STATES
View details

Peter K. Bode

-

SWITZERLAND
View details

Jack Shern

Head, Tumor Evolution and Genomics Section

National Institutes of Health

UNITED STATES
View details