This was my first course related to research, and also my first time working with medical imaging. When I heard that we would be doing independent research, I immediately realized that this course would undoubtedly be a great challenge for me. Independent research meant there was no clear “standard answer”; instead, I had to explore and persist on my own.
At the beginning of my ROP project, I was actually the first student in the class to finalize a research direction. I quickly chose skin tone bias in melanoma detection as my topic and decided to work with the ISIC dataset. At that time, I felt well prepared: even though I noticed that dark-skin samples were rare, I believed the number would be “enough.” I even imagined finishing the project in less than two months.
But soon, reality hit me. Out of more than 30,000 ISIC images, there were almost no dark-skin cases. After that, I kept switching datasets: PAD, Fitzpatrick17k, MSKCC. However, each of them had serious problems: some had almost no melanoma cases, some had almost no dark-skin samples, some images contained a lot of background noise rather than just lesions, and some lacked skin tone labels altogether. Even when I combined them, the total number of dark-skin melanoma images was barely more than one hundred. During that period, I felt like I was constantly “starting over,” and every time I thought I had found a breakthrough, it quickly fell apart.
In this struggle, I tried almost everything I could think of. I trained my own U-Net, experimented with CLIP, SVM, EfficientNet, and ResNet; I tested light-skin-trained models directly on dark-skin data; I even used YOLO to crop lesions in order to reduce background noise. My research focus also shifted again and again: from melanoma, to pigmented lesions, and finally to red scaly diseases; and my tasks shifted from classification to segmentation and back again. Altogether, I must have attempted more than a dozen different approaches, yet none of them produced satisfactory results.
As the deadline drew closer, my anxiety grew stronger. By the last month, despite all the models, tasks, and research objects I had tried, I still had no meaningful results to show. At times I felt completely lost, unsure of what else I could even do. In desperation, I wrote Dr. Tyrrell a very long email, confessing that I might not be able to continue and even considered abandoning the project altogether. I told him that if I could start over, I would never choose to study bias so hastily, but would first spend more time carefully understanding the limitations of the datasets.
That month was probably the hardest part of the entire ROP. I stayed up late almost every day, exhausted and anxious, sometimes even afraid to run my code because I expected yet another failure. Dr. Tyrrell was sometimes worried and even a bit frustrated, which made me feel sad, but I was also deeply grateful that he cared so much. In the final weeks, Giuseppe also began to support me more closely, and I truly appreciated his help. During that time, even the smallest result—no matter how unrepresentative—felt important enough for me to immediately share with Dr. Tyrrell and Giuseppe for feedback.
Finally, near the very end, something changed. About ten days before the deadline, I obtained a result that was still imperfect, but at least demonstrated a sign of bias. It was not a breakthrough, but it was enough to build a conclusion. In the last week, I focused on writing the report, experimenting with bias-mitigation methods, and managed to finish everything just in time.
Looking back on these four months, I went through so many emotions: the early excitement of being “ahead,” the anxiety of being overtaken, the regret and despair of repeated failures, and the relief of a small last-minute success. If you ask me what kept me going, I honestly don’t know, perhaps the support from Dr. Tyrrell and Giuseppe, perhaps the stubborn voice in my head saying “try one more time,” or perhaps just a little bit of luck.
Through this course, I developed a new understanding of medical imaging and machine learning: they are not only technical problems but also involve fairness, data limitations, and persistence throughout the research process. I realized that the true value of research is not in quickly achieving a perfect result, but in continuously experimenting, reflecting, and learning from failures. In the future, I hope to further explore fairness in medical imaging, especially to investigate why my findings differed from previous studies and how I can avoid or better explain such discrepancies. I believe this will not only help me improve my research methods but also allow me to move forward more confidently on my academic path.