TLDR: A research paper by Deborup Sanyal and Dr. Mingchen Gao introduces an AI model using DenseNet architecture to predict COVID-19 severity from lung CT scans. The study utilized the STOIC2021 dataset of 10,000 CT images, employing pre-processing techniques like resampling, cropping, and clipping. Comparing DenseNet-121 and DenseNet-169, the DenseNet-121 model achieved a higher accuracy of 75.58% in predicting COVID-19 severity, offering a promising tool to assist doctors in patient assessment and potentially reduce fatalities.
The global COVID-19 pandemic, which began in late 2019, presented unprecedented challenges to healthcare systems worldwide. With millions of lives lost, primarily due to respiratory system failure, the need for rapid and accurate diagnostic and prognostic tools became critical. A recent research paper explores the use of artificial intelligence, specifically deep learning, to predict the severity of COVID-19 infections based on Computed Tomography (CT) scans of the lungs.
Authored by Deborup Sanyal and Dr. Mingchen Gao from the Department of Computer Science and Engineering at the University at Buffalo, this study aims to assist medical professionals in assessing the severity of COVID-19 within one month of a positive test result. The core idea is to leverage the power of machine learning models, which are less prone to human error and can achieve high accuracy through extensive training on medical imaging datasets.
Leveraging CT Scans and Deep Learning
The researchers focused on using Convolutional Neural Networks (CNNs), a type of neural network particularly well-suited for image analysis. Unlike traditional algorithms that rely on pre-defined features, CNNs can automatically extract relevant features directly from raw image data, making them highly effective for complex tasks like medical image classification. This capability helps avoid complicated and expensive manual feature engineering.
For their project, the team utilized the STOIC2021 dataset, which comprises hundreds of CT scans from various patients, along with their reverse transcription–polymerase chain reaction (RT-PCR) test results, demographic information, and clinical symptoms. The CT scans, initially in .mha file format, were processed using the SimpleITK library. A substantial training set of 10,000 images and a validation set of 300 images were prepared.
Data Preparation is Key
Before feeding the images into the neural network, a crucial data pre-processing phase was undertaken. This involved several steps to transform the raw data into a more useful and efficient format:
-
Resampling: The original image dimensions of 512×512 pixels were reduced to 224×224 pixels. This reduction helps in saving storage space while maintaining the physical size and essential image information.
-
Cropping: This step focused on removing irrelevant noise or unwanted objects from the periphery of the images, ensuring that the model concentrates on the main subject – the lungs.
-
Clipping: By adjusting the contrast between the subject and the background, clipping enhanced the visual clarity of the images, making them more interpretable for the model.
DenseNet Architecture for Enhanced Prediction
The study employed DenseNet as its foundational architecture, specifically testing DenseNet-121 and DenseNet-169 models. DenseNet is known for its excellent accuracy with fewer parameters and its ability to improve gradient propagation by connecting every layer to every other preceding layer. This unique connectivity ensures that feature maps from earlier layers are directly accessible to later layers, promoting feature reuse and preventing the vanishing gradient problem often encountered in very deep networks.
The procedure involved reading and pre-processing the .mha files, extracting labels (probability of having COVID-19 and its severity), and combining them with the pre-processed images to create a PyTorch dataset. This dataset was then split into training and validation sets, managed efficiently using PyTorch’s DataLoader. The models were trained for 100 epochs using the Adam Optimizer with a learning rate of 0.01 and Pytorch.BCEWithLogitsLoss as the loss function.
Performance Insights
The experimental results showed a clear difference in performance between the two DenseNet architectures:
-
DenseNet-169: Achieved an accuracy of 63.25%.
-
DenseNet-121: Demonstrated a significantly higher accuracy of 75.58%.
The DenseNet-121 model, with its layers in its four dense blocks, proved to be more effective for this specific task compared to DenseNet-169, which has layers. The researchers noted that training the models on the entire dataset of 10,000 images for 100 epochs, as opposed to smaller subsets and fewer epochs in previous work, contributed to this improved accuracy.
Also Read:
- Deep Learning Advances Sarcopenia Diagnosis Using CT Imaging
- Advanced AI Generates Diverse Synthetic X-rays to Boost Diagnostic Accuracy and Fairness
Future Implications
While the 75.58% accuracy achieved by DenseNet-121 is not perfect, it provides a strong foundation for developing more advanced CNN and deep learning algorithms for COVID-19 detection and severity prediction. This research highlights the potential of AI to aid doctors and scientists in reducing COVID-19 related deaths by offering a reliable tool for early assessment. The authors also suggest that exploring other algorithms, such as MobileNet, could lead to even better results in the future.
For more detailed information, you can refer to the full research paper: COVID-19 PREDICTION BASED ON CT-SCANS OF LUNGS USING DENSENET ARCHITECTURE.
