VariViT: A Vision Transformer for Variable Image Sizes

Aswathi Varma; Suprosanna Shit; Chinmay Prabhakar; Daniel Scholz; Hongwei Bran Li; Bjoern Menze; Daniel Rueckert; Benedikt Wiestler

cs.CVFeb 16, 2026

VariViT: A Vision Transformer for Variable Image Sizes

Authors: Aswathi Varma, Suprosanna Shit, Chinmay Prabhakar, Daniel Scholz, Hongwei Bran Li, Bjoern Menze, Daniel Rueckert, Benedikt Wiestler

Organizations: Department of Neuroradiology, Technical University of Munich · Institute for Artificial Intelligence and Informatics in Medicine, Technical University of Munich · Department of Quantitative Biomedicine, University of Zurich · Athinoula A. Martinos Center for Biomedical Imaging, Harvard Medical School, Boston

Abstract

Vision Transformers (ViTs) have emerged as the state-of-the-art architecture in representation learning, leveraging self-attention mechanisms to excel in various tasks. ViTs split images into fixed-size patches, constraining them to a predefined size and necessitating pre-processing steps like resizing, padding, or cropping. This poses challenges in medical imaging, particularly with irregularly shaped structures like tumors. A fixed bounding box crop size produces input images with highly variable foreground-to-background ratios. Resizing medical images can degrade information and introduce artefacts, impacting diagnosis. Hence, tailoring variable-sized crops to regions of interest can enhance feature representation capabilities. Moreover, large images are computationally expensive, and smaller sizes risk information loss, presenting a computation-accuracy tradeoff. We propose VariViT, an improved ViT model crafted to handle variable image sizes while maintaining a consistent patch size. VariViT employs a novel positional embedding resizing scheme for a variable number of patches. We also implement a new batching strategy within VariViT to reduce computational complexity, resulting in faster training and inference times. In our evaluations on two 3D brain MRI datasets, VariViT surpasses vanilla ViTs and ResNet in glioma genotype prediction and brain tumor classification. It achieves F1-scores of 75.5% and 76.3%, respectively, learning more discriminative features. Our proposed batching strategy reduces computation time by up to 30% compared to conventional architectures. These findings underscore the efficacy of VariViT in image representation learning. Our code can be found here: https://github.com/Aswathi-Varma/varivit

Figures & tables

Figure 1: Selecting the optimal input crop size is essential for maximizing the representation quality (i.e., attention level) of ViTs . (a): 2D slice of the large tumor bounding box crop. (b): Attention map of the vanilla ViT model on the image, showing attention to background data rather than the tumor. (c): The image resized to a large fixed size shifts focus to distortions arising from the operation. (d): (Ours) Smaller image crop without resizing. In our method, attention is mainly given to desired tumor regions.

Figure 2: The VariViT model addresses the problem of handling images with different sizes by introducing a novel positional embedding resizing mechanism and employing different batching strategies. The model utilizes a fixed patch size, ensuring consistent patch embedding sizes across images while simultaneously adapting to different sequence lengths using a center and select resizing strategy.

Figure 3: (a): Tumor bounding boxes of two sizes are displayed, with the tumor center of mass aligned. (b): Our proposed center and select method for resizing position embeddings initializes fixed sinusoidal embeddings for the largest image size. Embeddings for other sizes are selected from the center of this grid. (c): We present two batching strategies: Custom Batch Sampler (CBS) with the same image sizes and Gradient Accumulation (GA) with varying image sizes within a batch.

Figure 4: t-SNE visualization of embedding layer output for IDH status classification. (a) : Vanilla ViT , (b) : Pix2Struct and (c) : VariViT-GA , all with an embedding dimension of 384. Notice the clearer separation of clusters in our model’s plot.

Panel of Table 1 — Table 1: Comparison of VariViT with baseline models for the IDH mutation status prediction task. Average training times visualized on the right (hours).

Panel of Table 2 — Table 2: Comparison of VariViT with baseline models for the primary vs. secondary brain tumor classification. Average training times are visualized on the right (hours).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: 3D Bounding Box Distribution of Glioma (a) and Brain Tumor (b) datasets. The x-axis represents the size of the 3D bounding box along the three dimensions, while the y-axis denotes the frequency of samples. The distribution of samples is divided into three equal bins based on the size of the bounding boxes.

Batching	Coordinates	AUC	F1-Score	MCC
CBS	Absolute	$\mathbf{0.933}\pm 0.013$	$\mathbf{0.712}\pm 0.032$	$\mathbf{0.646}\pm 0.036$
	Relative	$0.933\pm 0.009$	$0.697\pm 0.013$	$0.622\pm 0.019$
GA	Absolute	$\mathbf{0.945}\pm 0.007$	$\mathbf{0.744}\pm 0.036$	$\mathbf{0.684}\pm 0.034$
	Relative	$0.931\pm 0.016$	$0.718\pm 0.033$	$0.666\pm 0.031$

Appendix

Table 7: Comparison of absolute and relative position embeddings for both the batching methods on the glioma dataset. Note that the embeddings are learned, and resizing is done by interpolation.

Figure 6: 2D view of the cosine similarity visualization of VariViT ’s positional embedding for each of the three image sizes.

Figure 7: t-SNE Plots depicting VariViT CBS on the left and GA on the right, showcasing evaluations based on the highest scores across k-folds for Glioma (top) and Brain tumor (bottom) datasets.

Explore similar work

Feb 20, 2026

Effect of Patch Size on Fine-Tuning Vision Transformers in Two-Dimensional and Three-Dimensional Medical Image Classification

Vision Transformers (ViTs) and their variants have become state-of-the-art in many computer vision tasks and are widely used as backbones in large-scale vision and vision-language foundation models. While substantial research has focused on architectural improvements, the impact of patch size, a crucial initial design choice in ViTs, remains underexplored, particularly in medical domains where both two-dimensional (2D) and three-dimensional (3D) imaging modalities exist. In this study, using 12 medical imaging datasets from various imaging modalities (including seven 2D and five 3D datasets), we conduct a thorough evaluation of how different patch sizes affect ViT classification performance. Using a single graphical processing unit (GPU) and a range of patch sizes (1, 2, 4, 7, 14, 28), we fine-tune ViT models and observe consistent improvements in classification performance with smaller patch sizes (1, 2, and 4), which achieve the best results across nearly all datasets. More specifically, our results indicate improvements in balanced accuracy of up to 12.78% for 2D datasets (patch size 2 vs. 28) and up to 23.78% for 3D datasets (patch size 1 vs. 14), at the cost of increased computational expense. Moreover, by applying a straightforward ensemble strategy that fuses the predictions of the models trained with patch sizes 1, 2, and 4, we demonstrate a further boost in performance in most cases, especially for the 2D datasets. Our implementation is publicly available on GitHub: https://github.com/HealMaDe/MedViT

Massoud Dehghan, Ramona Woitek, Amirreza Mahbod

Research Center for Medical Image Analysis and Artificial Intelligence, Department of Medicine, Faculty of Medicine and Dentistry, Danube Private University, Krems an der Donau, 3500, Austria

Vision Transformer Medical Image Classification

Ask PDF

Feb 20, 2026

ZACH-ViT: Regime-Dependent Inductive Bias in Compact Vision Transformers for Medical Imaging

Vision Transformers rely on positional embeddings and class tokens encoding fixed spatial priors. While effective for natural images, these priors may be suboptimal when spatial layout is weakly informative, a frequent condition in medical imaging. We introduce ZACH-ViT (Zero-token Adaptive Compact Hierarchical Vision Transformer), a compact Vision Transformer that removes positional embeddings and the [CLS] token, achieving permutation-invariant patch processing via global average pooling. Zero-token denotes removal of the dedicated aggregation token and positional encodings. Patch tokens remain unchanged. Adaptive residual projections preserve training stability under strict parameter constraints. We evaluate ZACH-ViT across seven MedMNIST datasets under a strict few-shot protocol (50 samples/class, fixed hyperparameters, five seeds). Results reveal regime-dependent behavior: ZACH-ViT (0.25M parameters, trained from scratch) achieves strongest advantage on BloodMNIST and remains competitive on PathMNIST, while relative advantage decreases on datasets with stronger anatomical priors (OCTMNIST, OrganAMNIST), consistent with our hypothesis. Component and pooling ablations show positional support becomes mildly beneficial as spatial structure increases, whereas reintroducing a [CLS] token is consistently unfavorable. These findings support that architectural alignment with data structure can outweigh universal benchmark dominance. Despite minimal size and no pretraining, ZACH-ViT achieves competitive performance under data-scarce conditions, relevant for compact medical imaging and low-resource settings. Code: https://github.com/Bluesman79/ZACH-ViT

Athanasios Angelakis

BioML Lab, RI CODE, UniBw, Munich, Germany · EDS, Amsterdam UMC, Amsterdam, Netherlands

Vision Transformer Efficient ViTs Medical Imaging

Ask PDF

May 28, 2026

SwInception -- Local Attention Meets Convolutions

Sparse vision transformers have gained popularity as efficient encoders for medical volumetric segmentation, with Swin emerging as a prominent choice. Swin uses local attention to reduce complexity and yields excellent performance for many tasks but still tends to overfit on small datasets. To mitigate this weakness, we propose a novel architecture that further enhances Swin's inductive bias by introducing Inception blocks in the feed-forward layers. The introduction of these multi-branch convolutions enables more direct reasoning over local, multi-scale features within the transformer block. We have also modified the decoder layers in order to capture finer details using fewer parameters. We demonstrate a performance improvement on eleven different medical datasets through extensive experimentation. We specifically showcase advancements over the previous state-of-the-art backbones on benchmark challenges like the Medical Segmentation Decathlon and Beyond the Cranial Vault. By showing that the existing inductive bias in Swin can be further improved, our work presents a promising avenue for enhancing the capabilities of sparse vision transformers for both medical and natural image segmentation tasks. Code and pre-trained weights can be accessed at https://github.com/Eiphodos/SwInception.

David Hagerman, Roman Naeem, Jakob Lindqvist +3

Chalmers University of technology, 412 96 Gothenburg, Sweden · Zenseact, Lindholmspiren 2, 417 56 Gothenburg, Sweden

Hybrid CNN-Transformer Architectures Efficient ViTs Medical Image Segmentation 3D Medical Image Segmentation

Ask PDF