Self Supervised Low Dose Computed Tomography Image Denoising Using Invertible Network Exploiting Inter Slice Congruence

Sutanu Bera, Prabir Kumar BiswasWACV 2023

TL;DR

This paper introduces a novel self-supervised method for denoising low-dose CT (LDCT) images, a critical task for reducing patient radiation exposure without sacrificing image quality. By leveraging the inherent congruence between adjacent CT slices and an invertible neural network architecture, the method effectively denoises images without requiring paired clean and noisy data, outperforming existing unsupervised techniques.

What Problem Does This Paper Address?

Computed Tomography (CT) scans are vital for medical diagnosis, but higher radiation doses lead to better image quality. Low-dose CT (LDCT) aims to reduce this radiation exposure, but it comes at the cost of increased noise and artifacts, degrading diagnostic quality. Traditionally, deep learning models for image denoising require large datasets of paired noisy and clean images for training. However, obtaining such paired data in clinical settings is often infeasible. This paper addresses the challenge of training effective LDCT denoising models in a self-supervised manner, meaning without the need for corresponding clean (normal-dose CT) images.

Methods

The core of this work lies in two key innovations:

  • Inter-Slice Congruence: The authors exploit the fact that adjacent CT slices, while not identical, share significant structural information. They propose using the average of two neighboring slices as a proxy for a "second noisy observation" of the current slice. This is inspired by the Noise2Noise paradigm, which learns to denoise by mapping noisy inputs to other noisy versions of the same underlying clean image.
  • Invertible Neural Networks (INNs): To ensure that no information is lost during the denoising process and to act as a regularizer, the authors employ an invertible neural network. This architecture allows for a bijective mapping between input and output, meaning the transformation can be perfectly reversed. In their approach, the forward pass maps a noisy slice to the average of its neighbors (the denoised estimate), while the reverse pass maps this estimate back to the original noisy input. This reverse mapping acts as a cycle-consistency loss, preventing the network from losing crucial details or introducing artifacts, especially at organ boundaries.

The network architecture consists of encoders, decoders, and an invertible core. The forward mapping uses an encoder to process the noisy slice, an invertible core to transform the feature representation, and a decoder to produce the denoised image. The reverse mapping uses a similar structure but aims to reconstruct the original noisy input from the denoised output.

Datasets

The proposed method was evaluated on two datasets:

  • D1: 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge Dataset: This dataset contains simulated quarter-dose CT scans from 10 patients, with paired NDCT and LDCT images available. It was used for objective evaluation.
  • D2: ELCAP Public Lung Image Database: This dataset comprises real low-dose CT scans from 50 patients. It was used for subjective evaluation.

Key Insights and Ideas

  • Self-supervised denoising is feasible and effective for LDCT: The paper demonstrates that by cleverly leveraging inter-slice correlations, high-quality denoising can be achieved without paired ground truth data.
  • Invertible networks are powerful regularizers: The bijective nature of INNs provides a built-in mechanism to prevent information loss and enforce consistency, which is crucial for medical imaging where subtle details matter.
  • Cycle-consistency loss is vital for boundary preservation: The reverse mapping in the INN acts as a form of cycle-consistency loss, effectively mitigating artifacts and preserving organ boundaries that are often blurred or lost in other unsupervised methods.
  • Clinical deployment is simplified: The plug-and-play nature of the proposed method, which doesn't require paired data or interfere with existing CT scanner workflows, makes it highly practical for real-world clinical adoption.

Key Results

The proposed method (M3) demonstrated superior performance compared to other unsupervised methods (DIP, ConsensusNet, BM3D, CycleGAN) on the simulated dataset (D1). Specifically, it achieved the highest PSNR, SSIM, FID, and VIF scores. Visually, the method produced denoised images with less granular pattern and improved lesion visibility, comparable to the original NDCT images, while preserving texture. The reverse mapping also showed that the predicted noise pattern closely resembled the original noise pattern, validating the minimal information loss.

Limitations and Future Work

  • Limitations: While the paper shows excellent results, the authors acknowledge that the "impossible task" of mapping a noisy image to another noisy version means the loss (Lf) never fully minimizes. The resulting image is an estimate that minimizes deviation from noisy targets and is identical to the general class of deviation-minimizing estimators. The method's performance on real low-dose CT data is primarily evaluated qualitatively due to the absence of ground truth.
  • Future Work: The paper suggests that the proposed method is a straightforward plug-and-play solution that doesn't interfere with existing CT scanner workflows, implying its readiness for further clinical integration and testing. The success in improving lesion visibility also points towards potential applications in anomaly detection and diagnosis.

Summary updated

BibTeX
@inproceedings{bera2023self,
  author = {Sutanu Bera and Prabir Kumar Biswas},
  title = {{Self Supervised Low Dose Computed Tomography Image Denoising Using Invertible Network Exploiting Inter Slice Congruence}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
}
Figure