News

We propose the development of a high-efficiency foundation model tailored for the MedMNIST v2 benchmark, utilizing a novel architecture based on Multi-Resolution Tree-Structured Vector Quantization (TSVQ). While current foundation models often rely on computationally expensive transformers, our approach focuses on a hierarchical quantization strategy. By employing multi-resolution codebooks, we can effectively capture and represent both long-range structural dependencies and intricate, short-range local correlations inherent in diverse medical imaging modalities, from pathology slides to radiological scans.

The core innovation lies in the tree-structured organization of the latent space. Unlike flat codebooks used in traditional VQ-VAEs, TSVQ offers a logarithmic search complexity, significantly reducing the energy required for both training and inference. This alignment with “Green Learning” principles ensures that our model achieves state-of-the-art representation fidelity without the massive carbon footprint typically associated with large-scale AI. By optimizing the codebook search and minimizing redundant parameters, we aim to demonstrate that high-performance medical AI can be both sustainable and accessible on modest hardware.

This framework serves as a robust, domain-agnostic foundation. The learned representations are designed to be highly transferable, enabling the model to excel across a spectrum of downstream tasks. Crucially, this architecture addresses the “small data” problem in clinical medicine; by pre-training on the comprehensive MedMNIST suite, the model can be fine-tuned on smaller, domain-specific clinical datasets with superior accuracy and stability. Ultimately, we aim to expand this green learning paradigm to broader healthcare applications, empowering the medical community with scalable, low-power, and high-precision diagnostic tools.

Our work starts with a simple idea: the way different brain regions communicate can reveal a lot about what the brain is doing. Instead of treating EEG as a collection of separate channels, we view it as a network and study the connectivity patterns between regions. This kind of representation is useful because different brain states often produce different connectivity structures. In other words, brain connectivity maps can provide a more intuitive and informative picture of neural activity than raw signals alone.

In one of our studies, we use the direct Directed Transfer Function (dDTF) to build these maps. A key advantage of dDTF is that it captures not only whether two brain regions are related, but also the direction of information flow between them. This makes it a good tool for describing dynamic interactions in the brain. In particular, these connectivity patterns for mental workload show clear differences across conditions. As illustrated in Fig. 2, low and high workload states already present visibly different connectivity maps across several frequency bands, suggesting that they contain meaningful information for distinguishing cognitive states.

Based on this observation, we developed the framework shown in Fig. 1. We first decompose the EEG signals into multiple frequency bands and construct a connectivity map for each band. These multiband maps are then combined into a unified feature representation. From there, we progressively refine the features, selecting the most informative ones and transforming them into a more discriminative space before making the final prediction. In this way, our method leverages interpretable brain connectivity patterns while keeping the overall learning pipeline efficient, practical, and easy to extend to other EEG analysis tasks.

AI-driven medical imaging has emerged as a transformative force in modern healthcare, empowering clinicians to deliver more accurate, efficient, and personalized diagnostic and therapeutic strategies. By automatically analyzing CT, MRI, ultrasound, and other imaging modalities, AI systems can identify subtle patterns, precisely segment anatomical structures, and support clinical decision-making with enhanced accuracy and consistency. These advancements not only improve diagnostic performance but also streamline clinical workflows and ultimately elevate the overall quality of patient care.

Among the many tasks in medical image analysis, kidney and kidney tumor segmentation are pivotal for the management of renal diseases, particularly renal cell carcinoma. Precise delineation of kidneys and tumors is essential for quantitative tumor assessment, treatment planning, surgical navigation, postoperative monitoring, and radiomics research. Accurate segmentation enables clinicians to reliably estimate tumor burden, evaluate tumor–organ spatial relationships, and facilitate nephron-sparing surgical strategies, all of which directly influence patient outcomes. Given that manual segmentation is labor-intensive and prone to inter- and intra-observer variability, the development of automated, robust, and reliable segmentation methods has become increasingly critical for both routine clinical practice and large-scale research.

To address these limitations, we propose 3D-Cube Multi-Stage Green U-shaped Learning (GUSL), a novel multi-stage feed-forward machine learning framework for 3D medical image segmentation without backpropagation. GUSL is designed to be computationally efficient, interpretable, and environmentally sustainable, while maintaining competitive segmentation performance.

The proposed framework adopts a cascaded multi-stage segmentation strategy tailored to different anatomical tasks. As illustrated in Figure 1, distinct stages are designed for coarse-to-fine segmentation. First, the original CT volume is downsampled to a lower resolution, enabling efficient coarse localization of the kidney with reduced computational complexity. This low-resolution stage provides an approximate spatial position of the kidney, as highlighted in the red box. [...]

Histopathologic analysis is a key confirmatory step in the cancer diagnosis pipeline, where pathologists analyze a tissue section of interest for abnormalities and the extent of disease progression. The digitization of these tissue slides has enabled the use of AI for Whole Slide Image (WSI) Analysis, primarily to simplify pathologists’ laborious tasks and improve diagnostic accuracy. Due to the large size of these images, they’re split into smaller patches, and after analysis of individual patches, the results are aggregated to get the result for the WSI. Such a paradigm is called Multiple Instance Learning (MIL).

Architectural patterns surrounding the tumor regions are key indicators of angiogenesis and help in prognosis prediction. Architectural patterns are classified into 9 types and are grouped into three categories based on the underlying vasculature. While some patterns are often seen, others are rare and common only in higher-grade tumors. This causes a data imbalance that may affect training. To overcome this challenge, we propose an ensemble classifier for architectural pattern classification.

To capture local details and global context, we employ a multi-resolution feature encoder. At each resolution, the Saab transform is applied to obtain joint spatial-spectral representations. Representation learning is followed by a pooling operation to obtain compact representations. The pooled features from each resolution are concatenated to obtain a single feature vector, which is used to select the most discriminant features for the target classification task. An XGBoost binary classifier trained on these selected features predicts a confidence score for each architectural pattern. The confidence scores from multiple one-vs-one classifiers are aggregated to predict the architectural pattern in each patch.

One of the key components of Green Image Coding (GIC) is multi-grid control, which enables efficient and scalable bit allocation across the framework’s hierarchical layers. Unlike traditional hybrid codecs designed for single-layer encoding, GIC decomposes images into multiple hierarchical layers via resampling, referred to as a multi-grid representation. This decomposition effectively redistributes energy and reduces intra-layer content diversity, but it also creates a complex high-dimensional optimization challenge when attempting to allocate bits optimally across these various layers.

To make this problem tractable, we establish a theoretical foundation by defining the relationship between local and global rate-distortion (RD) models. We demonstrate that the global RD model can be derived from the local RD model of an individual layer by applying specific offsets to both rate and distortion. Notably, the distortion offset is a constant value determined by up-sampling processes and is unrelated to the compression process itself. This theoretical breakthrough reduces an intractable high-dimensional problem into a set of manageable sequential decisions.

Based on these findings, GIC implements a practical slope-matching-based rate control strategy. This strategy allocates bits across multiple grids by matching the slopes of consecutive RD curves. A primary advantage of this design is its modularity; the rate control module only requires information from two consecutive layers to function. This allows the module to be easily duplicated for any number of layers in the encoder, effectively decomposing the global rate-distortion optimization into a sequence of local optimizations to ensure a scalable balance between bit rate and image distortion.