Minimal Input, Maximal Impact: A Ct-Only Semi-Supervised Learning Pipeline for Lung Cancer Survival Prediction Across 12 Datasets Publisher



Salmanpour M R ; Falahati S ; Mousavi A ; Pouria A H ; Mehrnia S S ; Alizadeh M ; Gorji A ; Farsangi Z ; Safarian A ; Maghsudi M ; Rahmim A ; Yuan R
Source: Progress in Biomedical Optics and Imaging - Proceedings of SPIE Published:2026

Abstract

Accurate survival prediction in lung cancer is essential for personalized treatment. While PET/CT offers detailed staging, its cost and limited availability hinder scalability. CT remains the clinical standard for screening and monitoring, but automated segmentation often fails due to low contrast and tumor heterogeneity. We propose a semi-automated, clickpoint-guided framework that improves localization, enhances model focus, and reduces inter-observer variability - bridging clinical usability and deep learning (DL) accuracy. We developed a CT-only pipeline for predicting overall survival (OS) in lung cancer, utilizing 999 CT scans from 12 publicly available datasets, including independent external testing. Five DL segmentation models - 3D Attention U-Net, ReconNet, VNet, ResUNet, and SAM-Med3D - were trained on both whole-volume and ROI-cropped CT images. Radiomics features were subsequently extracted from segmentations and evaluated using 13-dimensional reduction and 6 classification algorithms under supervised (SL) and semi-supervised learning (SSL) frameworks. In SL, the labeled dataset was split into five folds for rotating training and validation. In SSL, four labeled folds trained a logistic regression to generate pseudo-labels for unlabeled data, which were combined with labeled data for retraining. Models were evaluated on the held-out fold and external test sets to assess the generalization benefit of unlabeled data. VNet network achieved the highest segmentation performance (Dice=0.83±0.07; IoU=0.71±0.09) using ROI-cropped images, significantly outperforming whole-volume inputs with Dice=0.21±0.06; IoU=0.15±0.04(paired t-test, p<0.01). In the OS prediction task, SSL, including Linear Discriminant Analysis and Autoencoder, radiomics features extracted from ReconNet - based masks, achieved an accuracy of 0.90±0.004, outperforming SL (accuracy=0.72±0.03; p<0.01). Notably, segmentations from our semi-automated pipeline led to better predictive performance than physician-delineated masks, likely due to consistent spatial focus and reduced inclusion of non-tumoral tissue. This CT-only, semi-automated pipeline enables robust, reproducible survival modeling using minimal input, offering a scalable, clinically viable solution ready for real-world deployment. © COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.