Week 5: Eddie Wei - Endoscope and Image Classification Model Updates
Multimodal Endoscope Design Updates
Figure 1: Imaging components housed within the distal end, with the addition of the bendable section, insertion cord, and endoscope handle to form the complete endoscope device
This week, I focused on updating the endoscope design to make it more visually intuitive and understandable for a broader range of clinicians. For these updates, I incorporated the endoscope handle for the user to hold, the insertion cord, and the bendable section, based on the current cystoscope used by my clinician (Figure 1). The key difference in this design is that the distal end houses components specifically designed for multimodal imaging using Optical Coherence Tomography and Confocal Fluorescence Microscopy. I also added an optical fiber port to the handle, which contains the optical fibers responsible for OCT and confocal imaging. These fibers can then be connected to the optical cart that houses the optical setup, which I have shown in previous blog reports. In addition, after dissecting the cystoscope used by Dr. Scherr, I incorporated a control lever into the handle design. This lever would move mechanical pull wires connected to the beginning of the bendable section, allowing the distal end to bend during use. Although further consideration is still needed to ensure that the optical fibers do not break or experience excessive strain during bending, this design could theoretically allow forward scanning to be performed while maintaining controlled maneuverability.
Figure 2: Layout of components housed in the distal end before entering the bendable tube
Current multimodal endoscopes that combine OCT and fluorescence microscopy are typically limited to either forward-viewing [1] or side-viewing imaging [2]. In contrast, my endoscope design accounts for the ability to switch between both imaging orientations interchangeably. One major concern is how to transition between forward scanning and side-viewing 360° scanning during a single procedure while minimizing additional sterilization steps and reducing the risk of contamination. The animations below address this challenge by demonstrating how additional distal-end components could be applied with a protective plastic sheath during operation, allowing the device to switch between forward-viewing and side-viewing 360° scanning. In addition, with a 2.4 mm diameter, as shown in Figure 2, this endoscope design could be applicable to multiple clinical domains. For urology specifically, the small diameter may allow the device to be inserted for imaging of the bladder wall and potentially advanced into the ureter toward the kidney, depending on the length of the insertion tube.
Imaging Classification Machine Learning Model Updates
Figure 4. Image classification test using the current Python CNN model. The test image of a Red Bull car on bottom right was classified as a car with 74% confidence.
Now that the endoscope design portion of my immersion project is complete, I have also been working on developing a convolutional neural network model in Python as a preliminary framework for image classification. At this stage, the model is trained using the CIFAR-10 dataset, which contains 10 image classes: planes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks. My current model uses three convolutional layers with rectified linear unit activation functions. The first convolutional layer applies 32 filters, while the second and third convolutional layers each apply 64 filters. Each convolutional layer is followed by a max-pooling operation to reduce the spatial dimensions while preserving important image features. The extracted features are then flattened and passed through a dense layer before reaching the final softmax layer, which outputs a likelihood estimate for each of the 10 possible classes. Although the current model is trained on general image data, the goal for immersion is to adapt this framework to classify OCT and confocal fluorescence microscopy images as benign, low-grade carcinoma, or high-grade carcinoma. For training data, my goal this week is to collect as many relevant OCT and confocal images as possible, then post-process them using GIMP so they can be fed into a new model based on my existing code. On Monday, with permission from Dr. Scherr, we plan to test imaging using his confocal histology scanner on extracted patient bladder samples. To address the current challenge of obtaining OCT imaging data, Dr. Scherr has reached out to the former CEO of Imalux, an OCT probe company, and I have also contacted the Department of Ophthalmology, from which I have received an initial response.
This week, I continued my clinical immersion experience by observing patient interviews and robotic urologic procedures. On June 29, I observed several patient interviews involving different urologic conditions. One case involved high-grade T1 carcinoma, where bladder removal was discussed as a possible treatment option to reduce the risk of disease progression. Another case involved a 43-year-old patient with a urinary mass and a left upper bowel mass. I also learned about a right kidney cyst that appeared concerning but was classified as Bosniak 3, meaning it required further clinical evaluation and monitoring even though the kidney itself appeared otherwise normal. On July 1, I observed an operating room case involving the right kidney and ureter. The procedure used a single-port robotic approach. The tumor mass on the right kidney, which had previously been identified on MRI near the abdomen and pelvis, was visualized and surgically addressed. This case helped me better understand how preoperative imaging is used to guide robotic surgery planning and intraoperative decision-making. On July 2, I observed another single-port robotic procedure involving removal of a diverticulum from part of the bladder, followed by surgical stitching to close and repair the area. This procedure was different from a typical robotic approach because the robot was inserted through the bladder rather than through the abdomen. A cystoscope was also used to help guide the surgical team to the correct region of the bladder and to ensure that the procedure did not interfere with nearby structures, including the right ureter and kidney.
Comments
Post a Comment