Deep Learning Based on CT Predicts Response to Chemoimmunotherapy in Lung Squamous Cell Carcinoma

Aging Dis 2024 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Predicting Whether Lung Cancer Will Respond to Chemoimmunotherapy Before Surgery

Clinical Context Neoadjuvant chemoimmunotherapy (NCI) - combining platinum-based chemotherapy with PD-1 inhibitors before surgery - has become a promising approach for resectable NSCLC. However, only a subset of patients achieve a major pathological response (MPR), defined as less than 10% residual viable tumor after treatment.

MPR as Surrogate Endpoint MPR predicts long-term outcomes: patients achieving MPR have significantly better progression-free survival and overall survival than non-responders. Identifying which patients will achieve MPR before giving NCI could spare non-responders its toxicity and direct them to alternative strategies.

The DL Model Researchers trained a ResNet50 convolutional neural network on contrast-enhanced CT images from 309 lung squamous cell carcinoma (LUSC) patients across three medical centers to predict MPR before treatment begins, achieving AUCs of 0.95 and 0.90 in two independent validation cohorts.

Beyond Prediction The study went further than performance metrics - it analyzed the genomic, immune, and microbiome correlates of predicted MPR versus predicted non-MPR, using the CT-based DL prediction as a surrogate to investigate the biological mechanisms underlying treatment response.

TL;DR: A ResNet50 deep learning model trained on pre-treatment contrast-enhanced CT images predicts major pathological response to neoadjuvant chemoimmunotherapy in LUSC with AUCs of 0.95 and 0.90 in two validation cohorts.
Pages 2-3
Architecture and Training Strategy for MPR Prediction

ResNet50 Architecture ResNet50 uses residual blocks with skip connections - the output of each block is the residual function F(x) plus the original input x. This prevents gradient vanishing in deep networks and enables efficient learning of subtle imaging features relevant to treatment response.

Three-Stage Training Training proceeded in three stages: first connecting a linear classifier to the frozen convolutional network for initial feature alignment; then connecting a fully connected network with 1,024 hidden nodes for classification; and finally unfreezing all parameters for end-to-end fine-tuning at a low learning rate.

CT Image Preprocessing Contrast-enhanced CT images were annotated with tumor bounding boxes, then cropped and resized for network input. Data augmentation (flips, rotations, brightness/contrast changes, elastic deformations) was applied exclusively to training images to expand the effective dataset.

MPR Definition and Labeling Pathological MPR was defined as less than 10% residual viable tumor in the resection specimen, assessed by two senior pathologists with over 10 years of experience who independently reviewed all specimens. Discrepancies were resolved through discussion.

TL;DR: ResNet50 with a three-stage training strategy and contrast-enhanced CT input was developed to predict MPR, with tumor bounding box cropping and aggressive data augmentation to handle the limited training set.
Pages 3-4
Validation Results Across Three Independent Cohorts

Discovery Cohort 200 patients formed the discovery cohort for model development. Of all 309 patients, 149 (48.22%) exhibited MPR after NCI, providing a near-balanced classification dataset that facilitated model training.

First Validation Cohort In the first independent validation cohort (60 patients), the ResNet50 model achieved an AUC of 0.95 (95% CI: 0.98-1.00) - exceptional performance indicating near-perfect discrimination between future MPR and non-MPR patients based on pre-treatment CT.

Second Validation Cohort In the second independent validation cohort (49 patients from a different institution), AUC was 0.90 (95% CI: 0.81-0.98) - strong and consistent performance across different imaging equipment and patient populations.

Independent Predictor Status Multivariate logistic regression confirmed the DL model score (low vs. high) as an independent predictor of MPR after adjusting for clinical variables including age, sex, stage, and clinical tumor characteristics.

TL;DR: The ResNet50 model achieved AUCs of 0.95 and 0.90 in two independent validation cohorts, with the DL score confirmed as an independent MPR predictor in multivariate logistic regression.
Pages 3-4
Understanding What the AI Is Looking At in the CT Image

Grad-CAM Principle Gradient-weighted Class Activation Mapping (Grad-CAM) generates visual explanations of neural network decisions by computing the gradient of the target class prediction with respect to the final convolutional layer's feature maps, producing a heatmap of image regions most important for the classification.

Salient Region Identification Applying Grad-CAM to predicted MPR cases revealed that the network focused predominantly on specific CT features within and around the tumor, including regions of inhomogeneous enhancement, central necrosis, and tumor-stroma interface characteristics.

Alignment with Radiology The heatmap patterns identified by Grad-CAM corresponded to features that experienced radiologists associate with treatment-sensitive tumors, validating that the model had learned clinically meaningful imaging patterns rather than spurious background correlations.

Black Box to Glass Box Grad-CAM transforms ResNet50 from an opaque black-box predictor into a partially interpretable tool - oncologists can review the highlighted regions to understand why the model predicted high or low MPR probability, building clinical trust and supporting radiologist oversight.

TL;DR: Grad-CAM visualization showed the model focused on biologically relevant tumor regions including inhomogeneous enhancement and necrosis, providing interpretable explanations aligned with radiological intuition.
Pages 4-5
Genomic, Immune, and Microbiome Correlates of DL-Predicted Response

Somatic Mutation Analysis TCGA LUSC genomic data was stratified by DL model classification (P-MPR vs. P-No-MPR). The P-MPR group showed specific somatic mutation patterns in four genes correlated with the imaging classification, revealing genomic underpinnings of the CT-visible features.

Gene Ontology and Pathways GSEA analysis identified that P-MPR tumors were enriched for gene ontology terms related to immune response and antigen processing/presentation - consistent with the hypothesis that immunologically hot tumors have distinct CT textures reflecting their immune infiltrate.

Tumor Microenvironment CIBERSORTx analysis of TCGA RNA-seq data showed significant differences in immune cell composition between P-MPR and P-No-MPR groups, with P-MPR tumors showing higher proportions of activated CD4+ and CD8+ T cells and lower proportions of suppressive cells.

Intra-Tumor Microbiome The presence of four distinct bacterial species varied significantly between P-MPR and P-No-MPR groups, suggesting that intra-tumor microorganisms contribute to the CT-detectable imaging phenotype and may influence immunotherapy response through local immune modulation.

TL;DR: DL-predicted MPR correlated with specific somatic mutations, enrichment of antigen processing pathways, higher CD8+ T cell infiltration, and distinct intra-tumor microbiome profiles - linking imaging features to multi-omic biology.
Pages 5-6
Moving from Proof-of-Concept to Clinical Deployment

Pre-Treatment Decision Support A validated DL MPR predictor could guide the NCI treatment decision: patients predicted to achieve MPR receive NCI followed by resection, while those predicted as non-responders could be enrolled in alternative neoadjuvant strategies or clinical trials.

LUSC-Specific Application LUSC was chosen as the study population partly because LUSC is less likely to harbor targeted mutation drivers (EGFR, ALK) than LUAD, making immunotherapy combination approaches more relevant. The DL model fills a specific gap where molecular biomarkers provide less guidance.

Extending to LUAD Future work should validate whether a similar ResNet50 approach predicts MPR in LUAD, where the biology and CT characteristics differ from LUSC, and where genomic biomarkers like EGFR mutation status play a stronger role.

Prospective Randomized Validation A prospective randomized trial in which DL-guided treatment selection (NCI for high DL score, alternative for low DL score) is compared to standard selection criteria would provide the clinical evidence needed to change practice guidelines and expand use of this non-invasive prediction tool.

TL;DR: The DL model could direct LUSC patients with predicted high MPR probability to NCI, while steering predicted non-responders toward alternatives; prospective randomized validation is the critical next step.
Citation: Open Access, 2024. Available at: PMC12096918.