AAI_2025_Capstone_Chronicles_Combined

3

our object detection task. Sample images of a negative and positive sample can be seen in Figure 1. These coordinate values were also scaled by half to match the image resizing. The patient ID is a unique identifier that was used to group images by patient. The slice number is the identifier for a specific image, or scan slice as it is referred to in computational tomography (National Institute of Biomedical Imaging and Bioengineering, 2025). The split variable is categorical with 0 for train, 1 for validation, and 2 for test.

The differences in original image sizes likely stemmed from variations in CT scanner hardware across multiple institutions contributing to the RSNA dataset. To combat data leakage within patients, we implemented a stratified group shuffle split, where we ensured that all slices from a single patient in our subset are contained within a single split while maintaining the class ratio in all splits. We also applied min-max normalization to scale pixel intensities to the range [0,1]. This handles the variations in brightness from the CT scan images, and also allows for faster model convergence. ​ Results from our exploratory data analysis (EDA) show patterns in the image data. We plotted a heatmap of bounding box centers, as seen in Figure 2, which shows a concentration in the center of the image. This suggests that our models should focus on the central region when predicting for bounding boxes. We also confirm in the EDA that pixel values are in the normalized range of [0,1], there are no overlapping patient IDs across splits, and that the image sizes are 256x256. Every positive sample contains a box and every negative sample does not. We can use this to set up our object detection models to get classification results to compare against.

Figure 1 Sample images of scan slices. Two negative samples on the top row. Two positive samples on the bottom row. During the subset creation process some data quality and structural issues were identified and handled. The original dataset is highly imbalanced, with significantly more non-fractured or negative slices than fractured positive slices. To combat this we employed undersampling to achieve a target ratio of 1 positive sample to every 3 negative samples. We decided to keep this ratio instead of an equal balance to keep the nature of the medical problem we are dealing with, as health events such as fractures are naturally rare compared to healthy anatomy (Gao et al., 2020). Another issue we handled was inconsistent image dimensions. To ensure consistency, we filtered out any images that were not 512x512 images, leaving us with 99% of the original dataset to subset and resize.

Figure 2 Heatmap of bounding box locations across all images.

306