Week 7: Eddie - Classification Model Updates

General Update

Unfortunately, Dr. Scherr was away this week, so I was unable to shadow any operations. However, the week was still very active, with additional presentations, discussions with SamanTree specialists and employees, and extensive coding. In addition to presenting on Tuesday, I was given two more opportunities to present my work to members of the SamanTree Histolog team and discuss the next steps for the final week of immersion. We are currently working to determine how the images acquired with the Histolog scanner can be exported and made available for my research. In the meantime, I identified confocal image data showing high-grade cancer, low-grade cancer, and healthy tissue in the inner lining of the bladder. Using a dataset of 48 images, I developed a preliminary computer model that examines each image, selects one of the three tissue categories, and provides a score indicating how strongly the image matches each category.

Model Changes and Features

Because the dataset I have access to right now is not large for more advanced learning methods, I made changes to my model that compares visible image characteristics. I selected and incorporated these characteristics into my code based on published research and patterns I observed during clinical procedures and while using the Histolog scanner. Throughout this experience, I observed that low-grade bladder cancer generally showed organized, densely packed cells of similar sizes and shapes, along with finger-like structures supported by blood-vessel stalks. High-grade cancer showed disorganized cells of varying sizes and shapes, unclear borders, reduced attachment between cells, and distorted blood vessels. Both cancer grades lacked normal umbrella cells. In contrast, normal urothelium contained large, polygon-shaped umbrella cells on the surface, with intermediate cells underneath. Based on these observations, I designed the model to examine many visible characteristics in each image. It measures brightness and contrast, including the distribution of light and dark areas, and determines whether the tissue appears smooth and consistent or rough and uneven. It also evaluates the number, size, shape, and spacing of bright tissue structures, including whether they are round, stretched, irregular, clustered, or scattered. For the largest visible structure, the model measures its length, width, thickness, boundary shape, and number of branches. It also examines enclosed openings, the amount and direction of visible edges, and how sharply brightness changes across the image. To detect small areas of abnormal tissue that might otherwise be overlooked, the model divides each image into smaller sections and compares their brightness, texture, sharpness, and irregularity. It also measures patterns appearing at different sizes and directions, including broad tissue structures and fine details that may not be easily recognized through visual inspection alone. Before selecting the current approach, I tested several classification methods, including a three-stage process that evaluated the tissue categories separately. However, these approaches produced poor and inconsistent results, particularly when distinguishing low-grade cancer from the other categories. After comparing the different methods, I found that a two-step process performed better. The model first determines whether an image is likely to show low-grade cancer. If it does not, the model then determines whether it shows high-grade cancer or normal tissue. I designed this approach to prioritize low-grade cancer because it has been the most challenging category for the model to identify reliably. To evaluate the model, I used a three-part testing process. The model was repeatedly trained on two portions of the dataset and tested on the remaining portion. The program reports how often its classifications are correct, how well it identifies each tissue category, and which categories it commonly confuses. It also displays each image alongside its actual category, predicted category, and matching scores for high-grade cancer, low-grade cancer, and normal tissue.

Results

Currently, my model achieved an overall accuracy of 54.17%, correctly classifying 26 of the 48 images. Fortunately, its strongest performance was on determining normal tissue, correctly identifying 12 of 16 normal images, or 75%. It also correctly identified 8 of 12 low-grade cancer images, or approximately 67%. Importantly, none of the low-grade images were incorrectly classified as normal; the four incorrect low-grade predictions were classified as high-grade. High-grade cancer remained the most difficult category. The model correctly identified 6 of 20 high-grade images, or 30%. Of the remaining high-grade images, nine were classified as low-grade and five as normal. These results suggest that the model is beginning to recognize patterns associated with normal and low-grade tissue but still has difficulty capturing the range of appearances found in high-grade cancer. The model also identified several image characteristics that were especially useful for detecting low-grade cancer. The most influential characteristic was whether bright tissue structures were concentrated within one area of the image or distributed more evenly. Other important characteristics included the consistency of tissue texture across different parts of the image, differences in edge density and sharpness between small image sections, and the number of grouped branches within the largest visible structure. Fine-scale repeating patterns also contributed to the model’s decisions. Overall, the results are encouraging for an early model developed from only 48 images, but not yet sufficient for dependable tissue classification. My next steps are to investigate why high-grade images are frequently confused with low-grade and normal tissue, refine and be on the look out for features used to describe high-grade cancer, and test the model on additional independent images. Access to the images collected with the Histolog scanner will be particularly important for determining whether the patterns identified from published images also appear in newly acquired tissue images. 

Next Steps and Key Takeways

Although the difficulty accessing data from the Histolog scanner due to institutional protocols and limited employee availability was frustrating at times and prevented me from pursuing the model I had originally planned, I now appreciate the experience. It pushed me to actually fully analyze the distinguishing features of low-grade, high-grade, and normal bladder tissue in greater depth while strengthening my ability to adapt. With Dr. Scherr’s approval, I have access to the current Histolog data and it’s more so looking into being able to physically use/export this data or move my code into a hospital desktop. My next step will be to test my existing model on these images and, if sufficient data are available, potentially use them to further train and improve the model next week. In parallel, I will finalize the cystoscope design, focusing primarily on fine-tuning the objective-lens configuration within the distal housing