CINR

Convolutional Implicit Neural Representation of pathology whole-slide images

DongEon Lee1, Chunsu Park2, SeonYeong Lee1, SiYeoul Lee1, MinWoo Kim2,3
1Department of Information Convergence Engineering, College of Information and Biomedical Convergence Engineering, Pusan National University, Yangsan, Korea, 2School of Biomedical Convergence Engineering, College of Information and Biomedical Engineering, Pusan National University, Yangsan, Korea 3Center for Artificial Intelligence Research, Pusan National University, Busan, Korea

Differences between original and reconstructed images


Abstract

This study explored the application of implicit neural representations (INRs) to enhance digital histopathological imaging.

Traditional imaging methods rely on discretizing the image space into grids, managed through a pyramid file structure to accommodate the large size of whole slide images (WSIs); however, the continuous mapping capability of INRs, utilizing a multi-layer perceptron (MLP) to encode images directly from coordinates, presents a transformative approach. This method promises to streamline WSI management by eliminating the need for down-sampled versions, allowing instantaneous access to any image region at the desired magnification, thereby optimizing memory usage and reducing data storage requirements.

Despite their potential, INRs face challenges in accurately representing high spatial frequency components that are pivotal in histopathology. To address this gap, we introduce a novel INR framework that integrates auxiliary convolutional neural networks (CNN) with a standard MLP model.

This dual-network approach not only facilitates pixel-level analysis, but also enhances the representation of local spatial variations, which is crucial for accurately rendering the complex patterns found in WSIs. Our experimental findings indicated a substantial improvement in the fidelity of histopathological image representation, as evidenced by a 3-6 dB increase in the peak signal-to-noise ratio compared to existing methods. This advancement underscores the potential of INRs to revolutionize digital histopathology, offering a pathway towards more efficient diagnostic imaging techniques.

CINR Model


Position Encoding

Multi-resolution Hash Grid Encoding sets up L 2D grids across the domain, with each grid representing a different level of resolution. It encodes each grid point via a hash table of size T, where each index contains F trainable parameters (features), represented as a feature vector.

Convolutional Implicit Neural Representation, CINR Model

The tensor position is processed using two parallel-network. The first flow, a standard MLP with two layers, focuses features of the target position, whereas the second flow, a CNN with and 3 × 3 kernels, encompasses the features from the surrounding target position. The outputs from these two flows are merged and passed additional CNN layer to integrate the features, culminating in the of the target values.





Fully Connected Layer :

Directly maps coordinates to pixel values without spatial context.


Convolution Layer :

Uses convolution to capture spatial relationships between pixels, effectively representing high-frequency details.

Result


Reconstructed Results - Qualitative evaluation

Reconstructed Results - Quantitative evaluation

CINR based Image zoom in

BibTeX


      @InProceedings{Lee_Convolutional_MICCAI2024,
        author = { Lee, DongEon and Park, Chunsu and Lee, SeonYeong and Lee, SiYeoul and Kim, MinWoo},
        title = { { Convolutional Implicit Neural Representation of pathology whole-slide images } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {pending}
}