AI-ADC: Channel and Spatial Attention-Based Contrastive Learning to Generate ADC Maps from T2W MRI for Prostate Cancer Detection

Kutsev Bengisu Ozyoruk, Stephanie A. Harmon, Nathan Lay, Enis C. Yılmaz, Ulaş Bağcı, Deborah E. Citrin, Bradford J. Wood, Peter Pinto, Peter L. Choyke, Baris TurkbeyJournal of Personalized Medicine 2024 · DOI

TL;DR

This paper introduces AI-ADC, a novel deep learning model that generates Apparent Diffusion Coefficient (ADC) maps from standard T2-weighted MRI scans. This approach bypasses the need for Diffusion-Weighted Imaging (DWI), potentially streamlining prostate cancer diagnosis and monitoring by reducing scan time and complexity.

What Problem Does This Paper Address?

Prostate cancer diagnosis and monitoring heavily rely on multiparametric MRI (mpMRI), which typically includes T2-weighted MRI (T2W MRI) and Diffusion-Weighted Imaging (DWI). DWI provides valuable information through Apparent Diffusion Coefficient (ADC) maps, which can indicate tumor aggressiveness and treatment response. However, acquiring high-quality ADC maps is challenging due to artifacts from patient motion and rectal gas, which can compromise diagnostic accuracy. Furthermore, the full mpMRI protocol is time-consuming (30-40 minutes), making it less accessible for high-throughput clinical settings. This paper aims to address these issues by developing a method to generate accurate ADC maps directly from T2W MRI, eliminating the need for DWI acquisition.

Methods

The core of the AI-ADC approach is a Generative Adversarial Network (GAN) architecture. Specifically, it employs a contrastive learning strategy to map T2W MRI to ADC maps. This involves training the model to recognize similar and dissimilar pairs between the two modalities, allowing it to learn subtle structural details.

Key methodological components include:

  • Channel and Spatial Attention Module (CBAM): Integrated into the generator, this module helps the network focus on relevant features within the prostate region, improving the sensitivity to abnormal areas.
  • Contrastive Loss Function: This loss encourages the model to learn robust latent space representations, preserving image content and improving the accuracy of the generated ADC maps. It utilizes a patch-wise noise contrastive estimation to measure similarity between T2W MRI patches and corresponding generated ADC patches.
  • Self-Regularization Loss: This component ensures that the generated ADC maps accurately reflect the prostate boundary and other anatomical structures.
  • ResNet-based Encoder and PatchGAN Discriminator: These standard deep learning components are utilized for the generator and discriminator, respectively.

The model is designed for unpaired image-to-image translation, meaning it doesn't require perfectly matched T2W MRI and ADC map pairs for training, which simplifies data preparation.

Datasets

The study utilized two primary datasets:

  • In-house NIH Dataset: This dataset comprised MRIs from 506 patients with suspected or known prostate cancer. For training and evaluation, 399 patients with acceptable image quality and 106 with high-quality images were used. A test set of 195 patients was also used. Both T2W MRI, ADC maps, and manually derived prostate segmentation masks were used.
  • ProstateX Dataset: This publicly available dataset was used for external validation. It included 347 examinations from 344 individuals. This dataset was collected between November 2016 and January 2017.

The paper also mentions an in-house prostate segmentation AI model that was used to generate segmentation masks, which were then compared to radiologist-annotated masks.

Key Insights and Ideas

  • Bypassing DWI: The most significant insight is the demonstration that high-quality ADC maps can be generated from T2W MRI alone, potentially eliminating the need for DWI acquisition. This could drastically reduce scan times and improve patient comfort.
  • Contrastive Learning for Detail Preservation: The use of contrastive learning proved crucial for capturing subtle variations and intricate structural details in the generated ADC maps, which is vital for accurate lesion detection and characterization.
  • Attention Mechanisms Enhance Focus: The integration of channel and spatial attention modules effectively guided the model to focus on tumor-specific regions within the prostate, leading to more relevant and accurate ADC map generation.
  • Segmentation Mask Importance: The study highlights that accurate prostate segmentation masks are critical for the AI-ADC model's performance. Without them, the model can struggle with variability in prostate size and generate less reliable outputs.

Key Results

The AI-ADC model demonstrated superior performance compared to state-of-the-art methods like CycleGAN, CUT, and StyTr²:

  • In-house Dataset (Radiologist Cropping):
    • Achieved a mean SSIM of 0.870 ± 0.065, outperforming others.
  • Achieved a lower FID of 30.656, indicating better image quality and realism.
  • Inference time was remarkably fast at 0.0065 ± 0.013 seconds per slice.
  • In-house Dataset (AI Segmentation Cropping):
    • Achieved a mean SSIM of 0.863 ± 0.068.
  • Achieved a lower FID of 31.992.
  • Inference time remained very low at 0.0062 ± 0.0126 seconds.
  • ProstateX External Validation Dataset:
    • Achieved a SSIM of 0.647 ± 0.045.
  • Achieved a lower FID of 113.876.
  • Inference time was 0.0044 ± 0.013 seconds.

Qualitatively, AI-ADC maps showed better delineation of prostate boundaries and intraprostatic zones, especially in cases with artifacts in the original ADC maps. The model also effectively highlighted hypointense lesions suspicious for cancer.

Limitations and Future Work

  • Single-Center Study and Limited Training Data: The primary training dataset was from a single hospital, which could limit generalizability. However, the promising results on the external ProstateX dataset mitigate this concern to some extent.
  • Need for Segmentation Masks: The AI-ADC model's performance relies on having accurate prostate segmentation masks, which can be time-consuming to obtain manually. The authors demonstrated that their in-house AI segmentation model can automate this process.
  • No Radiologist Interaction: The current study did not involve radiologist interaction for the AI-ADC output, which is a crucial step for clinical validation.
  • Lesion Detection Impact Not Formally Tested: While qualitative results suggest good lesion visualization, the direct impact of AI-ADC on lesion detection accuracy has not been formally evaluated.

Future work includes:

  • Multi-reader Studies: To assess the clinical adaptability and acceptance of AI-ADC by radiologists.
  • Enlarging Style Transfer Region: To reduce the dependency on precise segmentation masks.
  • Multimodal Approach: Integrating AI-ADC with other clinical indicators like PSA density for more comprehensive prediction of biopsy outcomes.
  • Formal Lesion Detection Studies: To quantify the model's impact on diagnostic performance.

Summary updated

Abstract

These results highlight the efficacy of our proposed model in generating ADC maps from T2W MRI, showcasing its potential for enhancing clinical diagnostics and radiological workflows.
BibTeX
@article{ozyoruk2024ai,
  author = {Kutsev Bengisu Ozyoruk and Stephanie A. Harmon and Nathan Lay and Enis C. Yılmaz and Ulaş Bağcı and Deborah E. Citrin and Bradford J. Wood and Peter Pinto and Peter L. Choyke and Baris Turkbey},
  title = {{AI-ADC: Channel and Spatial Attention-Based Contrastive Learning to Generate ADC Maps from T2W MRI for Prostate Cancer Detection}},
  journal = {Journal of Personalized Medicine},
  year = {2024},
  doi = {10.3390/jpm14101047},
}
Figure