There exist many different pulmonary nodules that can occur, both benign and malignant. Currently, a radiologist will detect and assign a probability of malignancy to these through CT-scans or X-rays, and then perform a biopsy to confirm the type. This procedure, however minor, is invasive and involves risks. But not doing a biopsy could be worse, since assigning risk to nodules strictly based on images can lead to significant disagreement between radiologists. This discrepancy can be clearly seen by the sheer amount of nodules that were only marked by a single radiologist in the highly referenced LIDC/IDRI database.
To circumvent this, DEMCON created an artificially generated dataset inspired by the style of the LIDC database. This synthetic dataset, instead of assigning a mere likelihood of malignancy, embeds the exact type of nodule as part of its annotation. Thus, a generated nodule is definitively categorized into one of 15 different types of pulmonary nodules.
All of these well-established setbacks led us to frame the following research question:
Currently, DEMCON is looking into using this synthetic data to classify pulmonary nodules utilizing Deep Learning. The primary objectives for this project phase were successfully addressed, as detailed below:
1. Automatic Detection & Classification We aimed to build a deep learning-based AI model for the manual detection and automatic classification of lung nodules, utilizing synthetic data to improve training efficiency and performance. The model is detailed by the architecture listed below:
2. Diminishing Decision-Making Opacity Using the detailed annotation information available in the synthetic dataset, we aimed to diminish the model's decision-making opacity (enhancing Explainable AI) by providing two key visual elements:
-
Nodule Reconstruction: Generating visual reconstructions of the subjected nodules.
-
Attention Heatmaps: Generating Grad-CAM heatmaps at both the parenchyma and nodule level, portraying exactly where the model focused most of its attention during decision making.
The synthetic data generated by DEMCON had some flaws, such as problems with the spatial coordinates, a limited amount of data, and seemingly missing nodules.
In spite of this setback, when evaluated on Top-3 Accuracy, the model yielded a respectable score which should prove highly useful as an assisting tool for pulmonologists. What is more, illustrations such as the attention heatmaps and reconstructed nodules turned out to be extremely useful elements, as they massively enrich the informational space available to the end-user (e.g., the pulmonologist) when assessing a sample’s category.