Seungho Jung

Spend fewer image tokens on simple backgrounds, then restore a dense grid for reconstruction.

1. 01Entropy-based quadtree patches
2. 02Compact ViT encoding
3. 03Mask-token grid restoration

Compute vs. dense ViT-B

17.5% less

Reconstruction PSNR

29.60 dB

STL-10 benchmark: 36.21 vs. 43.88 GFLOPS; PSNR 29.60 vs. 34.03 dB for dense ViT-B (Table 1, p. 8).

Trade-off: less computation, but lower reconstruction quality than the dense baseline and visible artifacts in some skipped regions.

Predict four-year knee osteoarthritis progression without assuming every patient follows the same path.

1. 01One baseline knee X-ray
2. 02Present and future latent spaces
3. 03Align stability; separate progression

Progression-risk AUC

0.7638

Prediction horizon

48 months

OAI cohort of 4,796 image pairs with Site D held out for validation/test. Compared baseline AUC: 0.7289 (Table I, p. 4).

The input is a baseline radiograph; future grades provide training supervision. Results describe the reported OAI evaluation.

Use image-level domain context to adapt SAM normalization for medical segmentation.

1. 01Extract image statistics
2. 02Interpret domain context
3. 03Condition adaptive normalization

OAI Dice, full model

66.33%

Gain over Med-SA

+3.79 pp

Noisy OAI ablation: Dice rises from 62.54% to 66.33%; IoU from 51.04% to 53.61% (Table II, p. 5).

This comparison isolates AdaLN and the domain interpreter. The separate denoised experiment reports different scores.

Trace anomalous sensor readings back to likely causes while accounting for changing, uncertain relationships.

1. 01Learn dynamic causal graphs
2. 02Calibrate forecast uncertainty
3. 03Rank residual and structural deviations

MSDS root-cause AC@1

0.827

SWaT root-cause AC@3

0.517

Mean AC@K on industrial benchmarks: MSDS AC@1 = 0.827 +/- 0.102; SWaT AC@3 = 0.517 +/- 0.037 (Table 3, p. 12).

AC@K measures recovery of root-cause variables among the top K candidates. The model does not lead every metric; RCG has higher SWaT AC@1.

Locate breast lesions and assign their class within one ultrasound-image model.

1. 01Ultrasound image features
2. 02MaskFormer mask classification
3. 03Lesion masks and class labels

Mean average precision

0.943

Malignant-lesion recall

0.900

Breast Ultrasound Image Dataset; reported 80/10/10 train/test/validation split. Results from Tables 4 and 5, pp. 12 and 14.

Illustrated cases include small, multiple and low-contrast lesions. The model predicts masks and normal, benign or malignant labels in a single framework.