Week 4: Eddie Wei - Image Classification Model

 

Figure 1: Preliminary Convolutional Neural Network Performance

Preliminary Image Classification Model

        With the multimodal OCT and confocal fluorescence microscopy endoscope design mostly complete, I have started shifting my focus toward developing an image classification model for bladder tissue samples. The long-term goal is to classify bladder samples as normal, low-grade, or high-grade in real time using OCT and confocal microscopy images as inputs. For my current progress, I have begun developing and programming my model in Python using VS Code and training it with imported image datasets from TensorFlow. At this stage, the program works under a convolutional neural network. As shown in the figure above, my code currently includes three convolutional layers that perform feature extraction and learning through convolution operations and rectified linear unit activation functions, followed by dimensionality reduction and final classification using the softmax equation. As shown, the current model achieved approximately 66% training accuracy after 10 training iterations.

        This result provides a useful starting point, but the main goal for the second half of immersion is to improve the model using actual bladder sample data. Ideally, data will be collected from the same bladder tissue regions using both confocal microscopy and OCT to support the multimodal imaging approach. Confocal microscopy data may be collected using the histology scanner, while OCT data would be collected from the same or closely matched tissue regions using an OCT probe. One current challenge is access to an OCT probe. Unfortunately, the OCT probe company previously used by Dr. Scherr has discontinued service, and he no longer has access to that probe. To address this, the immediate goal for next week is to identify and obtain access to an OCT probe as soon as possible, potentially through the departments of ophthalmology or dermatology. If anyone knows of, or has access to, a handheld OCT probe that may be available for research use, I would greatly appreciate it if you could reach out to me!

Patient Cystoscopies and Robotic Surgery Procedures

        On June 23rd, I shadowed standard patient cystoscopies and clinical interviews. During the first patient interview and imaging discussion, the patient presented with a malignant tumor in the left kidney. Surgical removal was discussed as a possible treatment option; however, the patient’s brain MRI showed findings concerning possible metastasis to the brain, which could affect treatment planning and prognosis. In addition, I observed several cystoscopy procedures. During one cystoscopy, three bladder findings were evaluated: two appeared normal, while one was identified as low-grade bladder cancer. Based on clinical experience, the low-grade lesion was treated by burning, or ablating, the abnormal tissue during the procedure. I learned that low-grade bladder cancer typically has a lower risk of spreading and may sometimes be treated during cystoscopy, while high-grade bladder cancer has a greater potential to spread and may require more aggressive surgical management. An additional attachment was used during the procedure to perform the ablation.

On June 24th, I observed three robotic surgery operations. The first robotic surgery involved removing a bladder tumor and identifying the underlying ureter, which had been covered by the bladder tissue. The purpose of this procedure was to allow the surgical team to carefully remove the tumor while locating and preserving the ureter. The second and third robotic surgeries involved complete prostate removal. The second case was observed by both Gabriella and I. On June 25th, the operation I observed first used robotic surgery to remove the bladder and parts of the small intestine. Then, after docking and removing the robot, standard operation was performed to stitch the small intestine and bladder, to form a neobladder, then place it back into the patient.