Competitions

The SIIM-ACR Pneumothorax Segmentation competition, hosted on Kaggle by the Society for Imaging Informatics in Medicine (SIIM) and the American College of Radiology (ACR), challenged participants to develop AI models capable of detecting and segmenting pneumothorax (collapsed lung) in chest X-ray images. Accurate identification of pneumothorax is critical for timely medical intervention, and this competition aimed to enhance diagnostic tools to assist radiologists and improve patient outcomes.

As a result of winning this competition, I was invited to speak at the 2019 C-MIMI conference in Austin, Texas, where I presented our solution to the medical imaging community.

The Home Credit Default Risk competition, hosted on Kaggle, invited participants to develop machine learning models to predict the repayment capabilities of clients using alternative data sources, such as telecommunications and transactional information. The goal was to enhance financial inclusion by enabling Home Credit to offer safe borrowing experiences to individuals with limited or no credit histories. Accurate predictions would ensure that clients capable of repayment are approved for loans with appropriate terms, while minimizing the risk of default.

The Avito Demand Prediction Challenge, hosted on Kaggle, tasked participants with predicting the demand for online classified ads on Avito, Russia's largest classifieds website. Competitors utilized various data modalities—including textual descriptions, images, and contextual information—to forecast the likelihood of an ad leading to a successful transaction. Accurate demand prediction is crucial for optimizing user experience and platform efficiency, helping sellers set appropriate expectations and pricing. This competition aimed to enhance Avito's ability to match supply and demand effectively, thereby improving overall user satisfaction.

The Freesound Audio Tagging 2019 competition, hosted on Kaggle by Freesound and Google's Machine Perception team, challenged participants to develop multi-label audio tagging systems capable of recognizing 80 diverse sound categories from real-world environments. Participants utilized both curated and noisy datasets, applying machine learning techniques to automatically tag audio clips. The primary evaluation metric was label-weighted label-ranking average precision (lwlrap), measuring the precision of ranked label lists for each test clip.

Intracranial hemorrhage, bleeding that occurs inside the cranium, is a serious health problem requiring rapid and often intensive medical treatment. For example, intracranial hemorrhages account for approximately 10% of strokes in the U.S., where stroke is the fifth-leading cause of death. Identifying the location and type of any hemorrhage present is a critical step in treating the patient. In this competition, hosted by the Radiological Society of North America (RSNA), participants were challenged to build an algorithm to detect acute intracranial hemorrhage and its subtypes from head CT scans, aiming to help reduce the time to diagnosis.

The competition required detecting multiple subtypes of hemorrhages, including epidural, intraparenchymal, intraventricular, subarachnoid, subdural, and any hemorrhage, making it a complex multi-label classification problem.

The iMaterialist (Fashion) 2019 competition, part of the Fine-Grained Visual Categorization Workshop (FGVC6) at CVPR, challenged participants to develop algorithms for accurate segmentation and attribute labeling of fashion images. Visual analysis of clothing has received increasing attention in recent years, with applications that could enhance shopping experiences for consumers and increase work efficiency for fashion professionals. This competition introduced a novel fine-grained segmentation task that unified both categorization and segmentation of rich and complete apparel attributes.

This competition represented an important step toward real-world applications in fashion AI, requiring models to not only identify clothing items but also understand their detailed attributes and precise pixel-level segmentation.