arXiv subjects

Mahmut S. Gokmen

Publications and source records attributed to Mahmut S. Gokmen.

9 recordsLinked to original sources

arXiv2026

DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CT

Chest CT is among the highest-volume imaging exams in medicine, yet expert voxel-level annotations are scarce and costly, motivating encoders that learn directly from unlabeled scans. We present DALE-CT, a family of 2D slice-based Vision Transformers trained from scratch on chest CT with the heuristics-free LeJEPA objective. We introduce depth-aware slab sampling, which draws self-supervised views from across a physical $z$-axis slab rather than a single slice, implicitly tasking the 2D encoder with representing how anatomy changes between neighboring slices. The frozen representations trace each scan as a smooth anatomical trajectory, recover cranio-caudal slice ordering without labels, and distinguish slices by the anatomy they contain rather than by position alone. This anatomical world model emerges without any 3D or positional supervision, and an otherwise-identical encoder trained on slices in isolation never develops it. Building on this backbone, we introduce dense auxiliary supervision into the pretraining objective, using anatomical and abnormality masks to supervise patch and slice tokens alongside the self-supervised loss, and we compare the resulting variants against a DINOv2 baseline continually pretrained on CT-RATE. Among nine public and in-house models evaluated under the same protocol, DALE-CT-2S is the strongest 2D model in-domain, reaching 0.825 Macro AUROC on CT-RATE, within 0.024 of COLIPRI-CRM and without any text supervision. We subsequently scale the supervision-free configuration to a $\sim$287k-scan multi-source pool, to our knowledge the largest reported chest-CT pretraining corpus. The resulting DALE-CT-0-L posts the best 2D external-transfer point estimates, and we release it as our recommended backbone with the full model family, training code, and evaluation pipeline.

cs.CV↗

arXiv2026

Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature

Self-supervised pretraining objectives are spatially uniform: the teacher temperature and the per-patch loss weight are identical everywhere in the image, so a lesion a few patches wide contributes no more to the training signal than the surrounding parenchyma. Prior work biases the views toward annotated regions, which changes what the model sees but adds no pressure on the objective. We instead condition the targets of self-distillation, a method we call SALT (Spatially Adaptive Label-guided Temperature). Weak, box-derived labels, available only during pretraining, define a compact region on the encoder's patch grid, inside which the teacher's softmax temperature is sharpened and the masked-patch loss is up-weighted. The objectives, the masking policy and the centering statistics are otherwise unchanged, and at every downstream use the encoder is a plain feature extractor with no labels and no conditioning. We evaluate by freezing the encoder and training only a lightweight multi-depth CenterNet-style head, detecting lesions in 3D on four CT cohorts, and we isolate the mechanism against a backbone identical in architecture, pretraining data, schedule and label-guided cropping but with no target conditioning. We report patch-level separability, 3D detection stratified by cohort and by lesion size, box quality, and a detector-free probe in which a single frozen patch embedding re-identifies a lesion in a follow-up scan without registration, masks or fine-tuning. Because the conditioning is expressed through a spatial indicator rather than through label semantics, the formulation admits any weak spatial annotation; we instantiate and validate it for lesions.

cs.CV↗

arXiv2026

Curriculum-Driven 3D CT Report Generation via Language-Free Visual Grafting and Zone-Constrained Compression

Automated radiology report generation from 3D computed tomography (CT) volumes is challenging due to extreme sequence lengths, severe class imbalance, and the tendency of large language models (LLMs) to ignore visual tokens in favor of linguistic priors. We present Ker-VLJEPA-3B, a four-phase curriculum learning framework for free-text report generation from thoracic CT volumes. A phased training curriculum progressively adapts a Llama 3.2 3B decoder to ground its output in visual features from a frozen, self-supervised encoder. Our visual backbone (LeJEPA ViT-Large) is trained via self-supervised joint-embedding prediction on unlabeled CTs, without text supervision. Unlike contrastive models (CLIP, BiomedCLIP), this language-free backbone yields modality-pure representations. Vision-language alignment is deferred to the curriculum's bridge and generation phases. This modality-agnostic design can integrate any self-supervised encoder into an LLM without paired text during foundation training. Methodological innovations include: (1) zone-constrained cross-attention compressing slice embeddings into 32 spatially-grounded visual tokens; (2) PCA whitening of anisotropic LLM embeddings; (3) a positive-findings-only strategy eliminating posterior collapse; (4) warm bridge initialization transferring projection weights; and (5) selective cross-attention freezing with elastic weight consolidation to prevent catastrophic forgetting. Evaluated on the CT-RATE benchmark (2,984 validation volumes, 18 classes), Ker-VLJEPA-3B achieves a macro F1 of 0.429, surpassing the state-of-the-art (U-VLM, macro F1 = 0.414) by 3.6%, and reaching 0.448 (+8.2%) with threshold optimization. Ablation studies confirm 56.6% of generation quality derives from patient-specific visual content. Code and weights are available.

cs.CV↗

arXiv2026

A Framework for Cross-Domain Generalization in Coronary Artery Calcium Scoring Across Gated and Non-Gated Computed Tomography

Coronary artery calcium (CAC) scoring is a key predictor of cardiovascular risk, but it relies on ECG-gated CT scans, restricting its use to specialized cardiac imaging settings. We introduce an automated framework for CAC detection and lesion-specific Agatston scoring that operates across both gated and non-gated CT scans. At its core is CARD-ViT, a self-supervised Vision Transformer trained exclusively on gated CT data using DINO. Without any non-gated training data, our framework achieves 0.707 accuracy and a Cohen's kappa of 0.528 on the Stanford non-gated dataset, matching models trained directly on non-gated scans. On gated test sets, the framework achieves 0.910 accuracy with Cohen's kappa scores of 0.871 and 0.874 across independent datasets, demonstrating robust risk stratification. These results demonstrate the feasibility of cross-domain CAC scoring from gated to non-gated domains, supporting scalable cardiovascular screening in routine chest imaging without additional scans or annotations.

cs.CV↗

arXiv2026

DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection

Coronary artery disease (CAD), one of the leading causes of mortality worldwide, necessitates effective risk assessment strategies, with coronary artery calcium (CAC) scoring via computed tomography (CT) being a key method for prevention. Traditional methods, primarily based on UNET architectures implemented on pre-built models, face challenges like the scarcity of annotated CT scans containing CAC and imbalanced datasets, leading to reduced performance in segmentation and scoring tasks. In this study, we address these limitations by introducing DINO-LG, a novel label-guided extension of DINO (self-distillation with no labels) that incorporates targeted augmentation on annotated calcified regions during self-supervised pre-training. Our three-stage pipeline integrates Vision Transformer (ViT-Base/8) feature extraction via DINO-LG trained on 914 CT scans comprising 700 gated and 214 non-gated acquisitions, linear classification to identify calcified slices, and U-NET segmentation for CAC quantification and Agatston scoring. DINO-LG achieved 89% sensitivity and 90% specificity for detecting CAC-containing CT slices, compared to standard DINO's 79% sensitivity and 77% specificity, reducing false-negative and false-positive rates by 49% and 57% respectively. The integrated system achieves 90% accuracy in CAC risk classification on 45 test patients, outperforming standalone U-NET segmentation (76% accuracy) while processing only the relevant subset of CT slices. This targeted approach enhances CAC scoring accuracy by feeding the UNET model with relevant slices, improving diagnostic precision while lowering healthcare costs by minimizing unnecessary tests and treatments.

eess.IV↗

arXiv2025

Magnification-Aware Distillation (MAD): A Self-Supervised Framework for Unified Representation Learning in Gigapixel Whole-Slide Images

Whole-slide images (WSIs) contain tissue information distributed across multiple magnification levels, yet most self-supervised methods treat these scales as independent views. This separation prevents models from learning representations that remain stable when resolution changes, a key requirement for practical neuropathology workflows. This study introduces Magnification-Aware Distillation (MAD), a self-supervised strategy that links low-magnification context with spatially aligned high-magnification detail, enabling the model to learn how coarse tissue structure relates to fine cellular patterns. The resulting foundation model, MAD-NP, is trained entirely through this cross-scale correspondence without annotations. A linear classifier trained only on 10x embeddings maintains 96.7% of its performance when applied to unseen 40x tiles, demonstrating strong resolution-invariant representation learning. Segmentation outputs remain consistent across magnifications, preserving anatomical boundaries and minimizing noise. These results highlight the feasibility of scalable, magnification-robust WSI analysis using a unified embedding space

eess.IV↗

arXiv2025

Vision Foundry: A System for Training Foundational Vision AI Models

Self-supervised learning (SSL) leverages vast unannotated medical datasets, yet steep technical barriers limit adoption by clinical researchers. We introduce Vision Foundry, a code-free, HIPAA-compliant platform that democratizes pre-training, adaptation, and deployment of foundational vision models. The system integrates the DINO-MX framework, abstracting distributed infrastructure complexities while implementing specialized strategies like Magnification-Aware Distillation (MAD) and Parameter-Efficient Fine-Tuning (PEFT). We validate the platform across domains, including neuropathology segmentation, lung cellularity estimation, and coronary calcium scoring. Our experiments demonstrate that models trained via Vision Foundry significantly outperform generic baselines in segmentation fidelity and regression accuracy, while exhibiting robust zero-shot generalization across imaging protocols. By bridging the gap between advanced representation learning and practical application, Vision Foundry enables domain experts to develop state-of-the-art clinical AI tools with minimal annotation overhead, shifting focus from engineering optimization to clinical discovery.

q-bio.QM↗

arXiv2024

Enhancing Low Dose Computed Tomography Images Using Consistency Training Techniques

Diffusion models have significant impact on wide range of generative tasks, especially on image inpainting and restoration. Although the improvements on aiming for decreasing number of function evaluations (NFE), the iterative results are still computationally expensive. Consistency models are as a new family of generative models, enable single-step sampling of high quality data without the need for adversarial training. In this paper, we introduce the beta noise distribution, which provides flexibility in adjusting noise levels. This is combined with a sinusoidal curriculum that enhances the learning of the trajectory between the noise distribution and the posterior distribution of interest, allowing High Noise Improved Consistency Training (HN-iCT) to be trained in a supervised fashion. Additionally, High Noise Improved Consistency Training with Image Condition (HN-iCT-CN) architecture is introduced, enables to take Low Dose images as a condition for extracting significant features by Weighted Attention Gates (WAG).Our results indicate that unconditional image generation using HN-iCT significantly outperforms basic CT and iCT training techniques with NFE=1 on the CIFAR10 and CelebA datasets. Moreover, our image-conditioned model demonstrates exceptional performance in enhancing low-dose (LD) CT scans.

eess.IV↗

arXiv2024

High Noise Scheduling is a Must

Consistency models possess high capabilities for image generation, advancing sampling steps to a single step through their advanced techniques. Current advancements move one step forward consistency training techniques and eliminates the limitation of distillation training. Even though the proposed curriculum and noise scheduling in improved training techniques yield better results than basic consistency models, it lacks well balanced noise distribution and its consistency between curriculum. In this study, it is investigated the balance between high and low noise levels in noise distribution and offered polynomial noise distribution to maintain the stability. This proposed polynomial noise distribution is also supported with a predefined Karras noises to prevent unique noise levels arises with Karras noise generation algorithm. Furthermore, by elimination of learned noisy steps with a curriculum based on sinusoidal function increase the performance of the model in denoising. To make a fair comparison with the latest released consistency model training techniques, experiments are conducted with same hyper-parameters except curriculum and noise distribution. The models utilized during experiments are determined with low depth to prove the robustness of our proposed technique. The results show that the polynomial noise distribution outperforms the model trained with log-normal noise distribution, yielding a 33.54 FID score after 100,000 training steps with constant discretization steps. Additionally, the implementation of a sinusoidal-based curriculum enhances denoising performance, resulting in a FID score of 30.48.

cs.LG↗