AI+X Global Talent Community

Ouch, My Knee Hurts

Yeshwanth — Fri, 04 Sep 2026 17:04:06 GMT

This project addresses the 2026 RSNA Knee Abnormality Detection AI Challenge, a Kaggle competition to detect twelve clinically important abnormalities on knee MRI. It is the first RSNA challenge to train models on medical images paired with free-text radiology reports, and the reports arrive in nine languages from sites worldwide. Our working set comprises 4,407 studies and 24,371 imaging series, evaluated as macro AUC across twelve findings spanning ligaments, menisci, cartilage, and soft tissue.

The defining constraint is supervision. Only 58 studies carry radiologist annotations; the remaining 4,349 provide only a report, and reports are present in the training split alone, so text can supply targets but can never be an input at inference. Our pipeline therefore translates and mines the reports into weak per-finding labels, recovers acquisition geometry and contrast from DICOM headers where the provided series metadata is degraded, and trains a vision transformer over multiple series per study with per-diagnosis attention over series slots.

Ouch, My Knee Hurts

Yun-Hsuan Ku — Thu, 27 Aug 2026 22:18:40 GMT

Ouch, My Knee Hurts!

Liu Bowen — Wed, 26 Aug 2026 20:27:32 GMT

Ouch, My Knee Hurts!

Jia Qi Choy — Wed, 26 Aug 2026 20:15:03 GMT

Chest X-ray Classification with ViT

Yuta Suzuki — Wed, 26 Aug 2026 19:24:44 GMT

This project develops a multi-label chest X-ray disease classification pipeline using the NIH ChestX-ray14 dataset. We identified class imbalance and patient-level data leakage as key challenges. To address them, we created patient-wise train/validation splits, reduced only “No Finding” images in the development set, and retained all disease-labelled images. We fine-tuned a pretrained DeiT-Tiny/16 model using BCEWithLogitsLoss with tempered positive weighting for rare diseases. On the test set, the selected model achieved a Macro-F1 of 0.275 and a Micro-F1 of 0.346. The results establish a baseline, but the increasing validation loss and lower rare-disease F1 scores indicate substantial overfitting and the need for further optimization.

Chest X-ray Classification with ViT

Kenta Ueda — Wed, 26 Aug 2026 19:22:25 GMT

This project develops a multi-label chest X-ray disease classification pipeline using the NIH ChestX-ray14 dataset. We identified class imbalance and patient-level data leakage as key challenges. To address them, we created patient-wise train/validation splits, reduced only “No Finding” images in the development set, and retained all disease-labelled images. We also matched disease prevalence between the training and validation sets, achieving a maximum prevalence gap of 0.73 percentage points across labels. We fine-tuned a pretrained DeiT-Tiny/16 model using BCEWithLogitsLoss with tempered positive weighting for rare diseases. On the unchanged official test set, the selected model achieved a Macro-F1 of 0.275 and a Micro-F1 of 0.346. Although the train and validation distributions were closely aligned, the increasing validation loss and lower rare-disease F1 scores indicate substantial overfitting and motivate further optimization.

Brain Tumour Segmentation and Volumetry with CNN-ViT

Monica Chiang — Wed, 26 Aug 2026 16:15:11 GMT

This project developed an end-to-end AI pipeline for automated brain tumour segmentation and volumetric analysis from multimodal MRI. Using the Medical Segmentation Decathlon BrainTumour dataset, we adapted an ImageNet-21k-pretrained hybrid CNN–Vision Transformer to four-channel MRI and evaluated it on a frozen, patient-level, tumour-volume-stratified split. The primary binary model achieved a mean 3D Dice of 0.8481 on 73 held-out patients, while a four-label extension explored tumour subregion segmentation. Beyond model evaluation, we translated the pipeline into a GPU-backed web application that accepts NIfTI scans, automatically performs segmentation and volumetry, provides interactive 2D/3D visualization, supports optional ground-truth comparison, and exports predicted NIfTI masks.

Diversity and the Student Rat Race: Can non-gameable criteria reduce competition over gameable ones?

Monica Chiang — Wed, 26 Aug 2026 16:00:44 GMT

This project studies strategic competition in university admissions through a broader question: can non-gameable criteria reduce competition over gameable ones? We model admissions as an equilibrium between universities and strategic students, distinguishing a manipulable academic signal from a non-manipulable diversity criterion. We compare different policy designs—including diversity weights, aggregate diversity targets, and protected pathways—and examine how each changes students’ incentives to invest in tutoring. Our main insight is that priorities, targets, and pathways are not equivalent: valuing or requiring diversity may redistribute or stratify the rat race, while protecting a non-gameable pathway can directly reduce incentives to game the academic signal.

Brain Tumour Segmentation and Volumetry with CNN-ViT

YUFEI CHANG — Wed, 26 Aug 2026 16:00:35 GMT

This project developed an end-to-end AI pipeline for automated brain tumour segmentation and volumetric analysis from multimodal MRI. Using the Medical Segmentation Decathlon BrainTumour dataset, we adapted an ImageNet-21k-pretrained hybrid CNN–Vision Transformer to four-channel MRI and evaluated it on a frozen, patient-level, tumour-volume-stratified split. The primary binary model achieved a mean 3D Dice of 0.8481 on 73 held-out patients, while a four-label extension explored tumour subregion segmentation. Beyond model evaluation, we translated the pipeline into a GPU-backed web application that accepts NIfTI scans, automatically performs segmentation and volumetry, provides interactive 2D/3D visualization, supports optional ground-truth comparison, and exports predicted NIfTI masks.

Diversity and the Student Rat Race: Can non-gameable criteria reduce competition over gameable ones?

Yun-Hsuan Ku — Wed, 26 Aug 2026 15:53:08 GMT