# Segmentation-Guided Hybrid Deep Learning for Pulmonary Nodule Detection and Risk Prediction from Multi-Cohort CT Images

## TL;DR

This paper introduces a deep learning framework that improves lung cancer screening by accurately detecting and predicting the risk of pulmonary nodules in CT scans.

## Contribution

A novel dual-stage framework combining segmentation-guided detection and a hybrid CNN–Transformer classifier for improved nodule localization and malignancy risk estimation.

## Key findings

- The framework achieved a CPM of 0.944 for nodule detection on the LUNA16 dataset.
- Malignancy classification on LIDC-IDRI reached a ROC-AUC of 0.988 and a PR-AUC of 0.947.
- Calibration metrics showed strong reliability with an expected calibration error of 0.209.

## Abstract

Background: Lung cancer screening using low-dose computed tomography (LDCT) demands not only early pulmonary nodule detection but also accurate estimation of malignancy risk. This remains challenging due to subtle nodule appearances, the large number of CT slices per scan, and variability in radiological interpretation. The objective of this study is to develop a unified computer-aided detection and diagnosis framework that improves both nodule localization and malignancy assessment while maintaining clinical reliability. Methods: We propose Seg-CADe-CADx, a dual-stage deep learning framework that integrates segmentation-guided detection and malignancy classification. In the first stage, a segmentation-guided detector with a lightweight 2.5D refinement head is employed to enhance nodule localization accuracy, particularly for small nodules with diameters of 6 mm or less. In the second stage, a hybrid 3D DenseNet–Swin Transformer classifier is used for malignancy prediction, incorporating probability calibration to improve the reliability of risk estimates. Results: The proposed framework was evaluated on established public benchmarks. On the LUNA16 dataset, the system achieved a competitive performance metric (CPM) of 0.944 for nodule detection. On the LIDC-IDRI dataset, the malignancy classification module achieved a ROC-AUC of 0.988, a PR-AUC of 0.947, and a specificity of 97.8% at 95% sensitivity. Calibration analysis further demonstrated strong agreement between predicted probabilities and true malignancy likelihoods, with an expected calibration error of 0.209 and a Brier score of 0.083. Conclusions: The results demonstrate that hybrid segmentation-guided CNN–Transformer architectures can effectively improve both diagnostic accuracy and clinical reliability in lung cancer screening. By combining precise nodule localization with calibrated malignancy risk estimation, the proposed framework offers a promising tool for supporting radiologists in LDCT-based lung cancer assessment.