Differentiation of Glioma Mimicking Encephalitis and Encephalitis Using Multiparametric MR-Based Deep Learning

Wenli Wu, Jiewen Li, Junyong Ye, Qi Wang, Wentao Zhang, Shengsheng XuFrontiers in Oncology 2021 · DOI

TL;DR

This paper introduces a deep learning approach using multiparametric MRI to differentiate between glioma mimicking encephalitis and encephalitis, two conditions with overlapping symptoms but vastly different treatments. The study demonstrates that pretrained Convolutional Neural Networks (CNNs), particularly Inception-v3 and ResNet-50, can accurately classify these conditions, outperforming human radiologists in diagnostic performance.

What Problem Does This Paper Address?

Accurately distinguishing between glioma and encephalitis is a critical challenge in neuroimaging. While both conditions can present with similar clinical symptoms and MRI findings, their treatment strategies and prognoses are drastically different. Lower-grade gliomas can sometimes mimic encephalitis, presenting with focal enhancement or no enhancement, leading to potential misdiagnosis and delayed treatment. Conversely, some forms of encephalitis can exhibit mass effects, leading to misdiagnosis as glioma. Traditional MRI interpretation relies on subjective visual assessment, which can be prone to error, especially in ambiguous cases. Advanced MRI techniques exist but are not always routinely available. This paper aims to leverage the power of deep learning and readily available multiparametric MRI data (FLAIR, CE-T1WI, T2WI) to develop an automated, accurate, and non-invasive diagnostic tool.

Methods

The core of this research lies in applying transfer learning with pretrained Convolutional Neural Networks (CNNs). The authors selected three popular CNN architectures: Alexnet, ResNet-50, and Inception-v3, which were initially trained on the massive ImageNet dataset for general image recognition. These pretrained models were then "fine-tuned" using a dataset of MRI scans from patients with glioma and encephalitis.

Key methodological aspects include:

  • Multiparametric MRI Input: The study utilized three standard MRI sequences: FLAIR (Fluid Attenuated Inversion Recovery), CE-T1WI (Contrast-Enhanced T1-Weighted Imaging), and T2WI (T2-Weighted Imaging).
  • Transfer Learning: Instead of training CNNs from scratch, which requires vast amounts of data, the researchers leveraged the feature extraction capabilities of models already trained on ImageNet. This significantly reduces the data requirements and training time.
  • Architecture Modification: To adapt the ImageNet-trained models (designed for 1000 categories) to a binary classification task (glioma vs. encephalitis), the fully connected layers were removed, and the output of the last convolutional layer was used as a deep feature representation. New, smaller fully connected layers were added for the final classification.
  • Data Augmentation: To further combat the limited dataset size, techniques like contrast transformation, brightness adjustment, sharpening, and flipping were employed to artificially increase the training data.
  • Evaluation Metrics: The performance of the models was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, and specificity. A five-fold cross-validation strategy was used to ensure robust evaluation.

Datasets

The study utilized a retrospective dataset comprising 3064 MRI brain images from 164 patients. This cohort was divided into two groups:

  • Glioma: 56 patients
  • Encephalitis: 108 patients

The dataset was split into training and testing sets. The authors did not introduce a new benchmark dataset but rather used their institutional data for training and evaluation.

Key Insights and Ideas

  • Deep Learning Outperforms Radiologists: A significant finding is that the deep learning models, particularly Inception-v3 and ResNet-50, achieved higher diagnostic performance (measured by AUC) than the two neuroradiologists who reviewed the same images. This highlights the potential of AI to augment or even surpass human diagnostic capabilities in specific tasks.
  • T2WI Modality is Highly Informative: For the best-performing models (Inception-v3 and ResNet-50), the T2WI sequence alone provided the highest classification accuracy. This is attributed to the characteristic hyperintense signal intensities seen in T2WI for both encephalitis and some gliomas, making it a strong indicator.
  • Model Architecture Matters: Inception-v3 and ResNet-50 consistently outperformed Alexnet. This is likely due to their more advanced architectures, with Inception-v3's "inception module" effectively capturing features at multiple scales and ResNet-50's "residual blocks" enabling deeper network training without degradation.
  • Single Modality Sufficiency: While multiparametric MRI is standard, the study found that single MRI modalities (FLAIR, CE-T1WI, or T2WI) could achieve high and similar classification performance when used with appropriate deep learning models. This is crucial for clinical translation, as it simplifies the workflow and reduces acquisition time.

Key Results

  • High Classification Performance: All three pretrained CNN models achieved AUC values over 0.9, indicating excellent performance.
  • Inception-v3 Achieves Highest Accuracy: The Inception-v3 model, using T2WI data, achieved the highest classification accuracy of 97.57%.
  • ResNet-50 and Inception-v3 Superiority: Inception-v3 performed statistically significantly better than Alexnet (p<0.05). While ResNet-50 and Inception-v3 showed no statistically significant difference between themselves, both were significantly better than Alexnet.
  • Radiologist Comparison: The AUC for the two neuroradiologists were 0.891 and 0.770. Inception-v3 and ResNet-50 showed significant differences compared to the radiologists' AUC (p<0.05), indicating superior performance. Alexnet's performance was comparable to an experienced radiologist.
  • T2WI Advantage: For Inception-v3 and ResNet-50, T2WI provided the highest accuracy, followed by CE-T1WI and FLAIR.

Limitations and Future Work

  • Small Sample Size and Single-Center Study: The authors acknowledge that the sample size was relatively small and the data came from a single institution. They suggest that multicenter datasets and larger patient cohorts are needed for further validation.
  • Exclusion of T1WI: The study did not evaluate the diagnostic performance of deep learning based on T1WI sequences alone, as it was difficult to automatically identify lesions in this modality.
  • Lack of Clinical Data Integration: The study did not incorporate clinical data or imaging features into the deep learning models. Future work could involve training clinical models that combine imaging and clinical information for potentially enhanced diagnostic accuracy.
  • Combined Sequences: While single modalities performed well, the authors note that combining multiple MRI parameters can often lead to a better understanding of tumor characteristics and enhanced classification performance. Future research could explore optimal combinations of sequences.

Summary updated

Abstract

The pretrained CNN models can automatically and accurately classify these two diseases and further help to improving clinical diagnostic performance.
BibTeX
@inproceedings{wu2021differentiation,
  author = {Wenli Wu and Jiewen Li and Junyong Ye and Qi Wang and Wentao Zhang and Shengsheng Xu},
  title = {{Differentiation of Glioma Mimicking Encephalitis and Encephalitis Using Multiparametric MR-Based Deep Learning}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2021.639062},
}
Figure