Week 8 - Eddie Wei: Final Image Classification AI Model Results and Ending Thoughts

Bladder Image Collection and Classification Model Explanation 

              Figure 1: Some screenshots of bladder images segmented and classified (81 Images Total)

During the final week of the immersion, I was able to coordinate access to the bladder images collected with the Histolog system and begin using them for analysis. Using ImageJ, I selected and segmented regions representing high-grade, low-grade, and normal bladder tissue. These classifications were confirmed in collaboration with the pathology department using corresponding H&E-stained sections and discussions. In addition, during a prostate removal multiport robotic surgery this Wednesday, Dr.Scherr was able to extract additional normal bladder tissue that I was missing and imaged them in the operating room with the SamanTree histolog scanner. From the bladder specimens imaged over the previous two weeks, I generated 28 high-grade, 22 low-grade, and 28 normal tissue images. Each image was standardized to 500 × 500 pixels and converted to 8-bit grayscale for analysis. As for my model, its a compact two-stage Random Forest classifier to distinguish low-grade, high-grade, and normal bladder tissue to accommodate for the sample size of data. During training and testing, the model extracts 14 quantitative features from each image, including dark-to-light ratio, a local nuclear-to-cytoplasm ratio proxy, variation in nuclear-like size and shape, direction variation, crowding, intensity variation, texture homogeneity, spacing, contour irregularity, mean object size, and local dark-density variation. These features were selected based on reading Chapter 2 of Precision Molecular Pathology of Bladder [1], as well as discussions with Dr. Robertson about the characteristics used to distinguish high-grade, low-grade, and normal bladder tissue, as were segmenting areas on the histolog scanner.

The first stage identified low-grade tissue versus all other images. Images not classified as low grade were then evaluated by a second model that distinguished high-grade from normal tissue. Each stage used 200 decision trees that learned their own feature cutoffs from the labeled training images rather than relying on manually assigned thresholds. Because the bladder samples were collected from only three patients on different days, I used three-fold grouped cross-validation, keeping all images from the same patient together to reduce data leakage. This provides a more realistic estimate of performance on unseen patients. 

Model Demonstration and Results

                                     Figure 2: Quantitative performance metrics of model                                                                                   

As shown above, the final model achieved 82.72% accuracy, correctly classifying 67 of 81 images. It correctly identified 22 of 28 high-grade images, with 3 classified as low grade and 3 as normal; 21 of 25 low-grade images, with 4 classified as high grade; and 24 of 28 normal images, with 4 classified as low grade. No normal images were classified as high grade, and no low-grade images were classified as normal. Overall, the results were balanced across all three classes, but because the dataset came from only three patients, the findings remain preliminary and should be validated on additional independent specimens.

Future Steps 

Going forward, the next steps would be to collect additional confocal images and potentially acquire co-registered Optical Coherence Tomography images from the same regions to provide depth-resolved information. Further work could also focus on verifying and testing the multimodal OCT–confocal cystoscope design that I developed in SolidWorks and simulated in Zemax OpticStudio. The system is intended to support forward-facing imaging of the bladder using both modalities, while also enabling 360-degree side scanning with a 45-degree microreflector to image the ureteral and urethral linings leading to the bladder.

Acknowledgements

I would like to thank Phil Maher and Alison Gorman, from SamanTree, for taking time from their busy schedules to make confocal imaging of the bladder samples possible with their histolog scanner and dealing with me - sorry Alison for making you miss the train yesterday :(. I also appreciate Dr. Robinson in the Department of Pathology for taking the time this Wednesday and yesterday to review the samples and corresponding H&E stains to help confirm regions of low-grade, high-grade, and normal bladder tissue for me to segment and crop. I would like to acknowledge Dr. Scherr for giving me the opportunity to learn more about surgery, observe how medical professionals explain complex concepts to patients, and helping me develop the biological intuition that will shift my mindset in multimodal imaging with OCT and ultrasound. Most importantly, a huge thank-you to Kelly and Carly for organizing the immersion and making this entire experience something I will never forget.

References

[1] Robinson, B. D., & Khani, F. (2018). Grading, staging, and morphologic risk stratification of bladder cancer. In D. E. Hansel & S. P. Lerner (Eds.), Precision molecular pathology of bladder cancer (pp. 29–42). Springer. https://doi.org/10.1007/978-3-319-64769-2_2