The performance of vision models in medical imaging is often hindered by the prevailing paradigm of fine-tuning backbones pre-trained on out-of-domain natural images. To address this fundamental domain gap, we propose MedDChest, a new foundational Vision Transformer (ViT) model optimized specifically for thoracic imaging. We pre-trained MedDChest from scratch on a massive, curated, multimodal dataset of over 1.2 million images, encompassing different modalities including Chest X-ray and Computed Tomography (CT) compiled from 10 public sources. A core technical contribution of our work is Guided Random Resized Crops, a novel content-aware data augmentation strategy that biases sampling towards anatomically relevant regions, overcoming the inefficiency of standard cropping techniques on medical scans. We validate our model's effectiveness by fine-tuning it on a diverse set of downstream diagnostic tasks. Comprehensive experiments empirically demonstrate that MedDChest significantly outperforms strong, publicly available ImageNet-pretrained models. By establishing the superiority of large-scale, in-domain pre-training combined with domain-specific data augmentation, MedDChest provides a powerful and robust feature extractor that serves as a significantly better starting point for a wide array of thoracic diagnostic tasks. The model weights will be made publicly available to foster future research and applications.
Deploying transformer-based single object trackers on resource-constrained devices remains challenging because adapting a pre-trained backbone conventionally requires fine-tuning on the order of 90M parameters. Parameter-Efficient Fine-Tuning (PEFT) mitigates this cost, yet existing PEFT trackers often trade tracking accuracy for efficiency. We introduce SVDTrack, a PEFT method that uses the Singular Value Decomposition (SVD) of the pre-trained attention and MLP weight matrices to decide where adaptation should occur. SVDTrack combines two targeted modifications: (1) a learnable reweighting of the singular values, S = w ⊙ Σ + b, and (2) low-rank adaptation restricted to the singular vectors associated with the least significant singular values. We show that any update confined to this residual singular subspace provably preserves the optimal rank-k approximation of the pre-trained weights—the model’s core knowledge—and we introduce a constrained Residual LoRA variant that satisfies this condition by construction, which is essential when adapting weights already fine-tuned on tracking data. After training, all learned modifications are reparameterized back into standard weight matrices, so the deployed model incurs zero additional inference latency or memory. On four benchmarks (LaSOT, LaSOText, TrackingNet, GOT-10k), SVDTrack sets a new state of the art among PEFT trackers and matches or surpasses fully fine-tuned models on several metrics, while training only 9.57M parameters—an 89.6% reduction relative to full fine-tuning. Controlled ablations show that adapting the residual subspace outperforms adapting the most significant or randomly selected components under a matched parameter budget, supporting principled, SVD-guided adaptation as a practical route to efficient, high-performance tracking on edge devices.
3 November 2026 in Computer Vision and Image Understanding