Introduction
Lung cancer remains a primary cause of global mortality, making early detection through Computed Tomography (CT) scans a clinical priority. While Computer-Aided Diagnosis (CAD) systems have evolved, they often present a trade-off between the transparency of hand-engineered radiomics and the high performance of "black box" deep learning models. This paper introduces Stacked Interpretable Sequencing Cells (SISC), an architecture designed to bridge this interpretability gap by providing high-accuracy predictions alongside visual "critical response maps" that explain the model's reasoning.
The SISC Architecture
The SISC model utilizes a modular design based on repeatable "cells" rather than a monolithic network. Each sequencing module consists of three convolutional layers paired with batch normalization, dropout (0.25) for regularization, and ReLU activation functions.
As data progresses through the stack, the number of channels increases while kernel sizes remain fixed, simplifying the design process. The final stage employs a Global Average Pooling (GAP) layer and a Softmax activation to produce malignancy probabilities.
Dataset and Experimental Setup
The researchers evaluated SISC using the LIDC-IDRI dataset, comprising 1,018 thoracic CT scans. From these, 2,669 nodules $\ge 3$mm were extracted. To address the ambiguity in radiologist malignancy scores (1–5), three dataset variants were created:
- Dataset I (Ignored): Score 3 excluded.
- Dataset B (Benign): Score 3 treated as benign.
- Dataset M (Malignant): Score 3 treated as malignant.
Data augmentation techniques, including rotations and flips, were applied to expand the training set to approximately 15,000 samples per dataset, ensuring robust model training.
Interpretability and Results
The core innovation of SISC is its interpretability mechanism. By utilizing a deconvolution-based approach, the model backpropagates its final decision to the input layer, highlighting the specific pixels that influenced the classification.
Quantitative Performance
SISC achieved state-of-the-art results across all configurations:
- Dataset I: 89.36% accuracy and 96.01% AUC.
- Dataset B: 88.57% accuracy and 94.34% AUC.
- Dataset M: 84.17% accuracy and 89.06% AUC.
Qualitative Insights
The generated critical response maps frequently highlighted nodule contours and internal textures—features traditionally used by radiologists. This alignment suggests that the model "discovered" clinically relevant radiomic features autonomously. In many cases, the SISC-generated maps provided more precise nodule localization than manual radiologist masks.
Conclusion
SISC demonstrates that deep learning models for medical imaging do not have to be "black boxes." By combining a modular, high-performance architecture with a robust interpretability mechanism, SISC provides the "evidence-based AI" necessary for clinical adoption. Future work will focus on extending this architecture to 3D volumetric data and integrating automated nodule localization into the end-to-end pipeline.
Summary updated