Application of Deep Learning in Automated Analysis of Molecular Images in Cancer: A Survey

Yong Xue, Shihui Chen, Jing Qin, Yong Liu, Bingsheng Huang, Hanwei ChenContrast 2017 · DOI

TL;DR

This paper surveys the burgeoning field of applying deep learning (DL) to molecular imaging for cancer analysis. It highlights how DL, particularly Convolutional Neural Networks (CNNs), is revolutionizing tumor segmentation, classification, and survival prediction by overcoming the limitations of manual interpretation and traditional machine learning. The review emphasizes DL's potential to improve early cancer detection and personalized treatment.

What Problem Does This Paper Address?

Cancer remains a leading cause of death, and early and precise diagnosis is crucial for improving patient outcomes. Molecular imaging techniques offer a powerful way to visualize and quantify biological processes at the molecular level, often detecting cancer changes before they are apparent in anatomical imaging. However, analyzing these complex images is challenging for human experts. Visual assessment by radiologists is time-consuming, prone to inter-observer variability, and can miss subtle changes due to low signal-to-noise ratios. Traditional machine learning methods also struggle, often requiring manual feature extraction, which can be subjective and miss important information. This paper addresses the need for automated, accurate, and efficient analysis of molecular cancer images, exploring how deep learning can provide a solution.

Methods

The paper focuses on the application of deep learning techniques, with a strong emphasis on Convolutional Neural Networks (CNNs). CNNs are highlighted for their ability to automatically learn hierarchical features directly from data, overcoming the limitations of manual feature engineering. The review discusses various CNN architectures, including:

  • 2D CNNs: Applied for brain tumor segmentation, often by processing images slice-by-slice or fusing information from multiple orthogonal planes.
  • 3D CNNs: Used for more comprehensive volumetric analysis, capturing spatial context more effectively.
  • Ensemble methods: Combining multiple CNNs or CNNs with other techniques to improve robustness and performance.
  • Autoencoders: Employed for feature learning and dimensionality reduction, particularly in semi-supervised or weakly supervised learning scenarios.

The paper also touches upon other deep learning approaches like stacked denoising autoencoders (SDAE) and the use of conditional random fields (CRF) for post-processing and refining segmentation results. The core idea is that deep learning models can learn complex, discriminative features that are more effective for cancer image analysis than traditional methods.

Datasets

The paper references several datasets commonly used in cancer molecular imaging research:

  • BRATS (Brain Tumor Segmentation) dataset: A widely used benchmark for brain tumor segmentation, providing multimodal MRI sequences (T1W, T1-postcontrast, T2W, and FLAIR). The paper notes that BRATS has been divided into different classes and tumor regions since 2013.
  • Diffusion-weighted magnetic resonance images (DW-MRI): Used for prostate cancer classification.
  • Fluorodeoxyglucose (FDG) PET and 11C-methionine (MET) PET images: Utilized for brain tumor grading and classification.
  • Contrast-enhanced CT (CECT) images: Employed for survival prediction in lung cancer.

The paper does not introduce any new benchmark datasets but relies on existing public challenges and datasets for evaluating the discussed methods.

Key Insights and Ideas

  • Deep Learning's Superior Feature Extraction: DL models, especially CNNs, can automatically learn rich, hierarchical features from raw image data, eliminating the need for manual, potentially biased, feature engineering. This is a significant advantage over traditional machine learning.
  • End-to-End Learning: DL models can often perform feature extraction, selection, and classification in a single, integrated pipeline, leading to more optimized and accurate results.
  • 3D CNNs for Enhanced Spatial Context: While 2D CNNs are computationally less demanding, 3D CNNs offer superior performance for tasks like tumor segmentation by capturing the full volumetric context of the lesion.
  • The Power of Multimodal Imaging: Integrating information from different imaging modalities (e.g., multiple MRI sequences, PET/CT) through DL can significantly improve diagnostic accuracy.
  • Transfer Learning as a Solution for Data Scarcity: When training data is limited, pre-training models on large datasets (like ImageNet) and then fine-tuning them for specific medical imaging tasks can significantly boost performance.
  • Challenges in Data Imbalance: Datasets often suffer from class imbalance (e.g., rare tumor types or specific lesion characteristics), which can hinder DL model performance, particularly in segmentation tasks.

Key Results

The paper summarizes results from various studies, demonstrating the effectiveness of deep learning in cancer molecular imaging:

  • Tumor Segmentation: CNN-based methods have achieved state-of-the-art results on benchmarks like BRATS, with reported accuracies and Dice similarity coefficients often exceeding 0.85, and in some cases, reaching over 0.99. Some methods also achieved significant speedups (e.g., 40-fold faster) compared to previous approaches.
  • Cancer Classification: DL models have shown high accuracy in classifying various cancers, such as prostate cancer (e.g., accuracy of 0.943) and brain tumors.
  • Survival Prediction: Pretrained CNN architectures have demonstrated high accuracy in predicting patient survival times (e.g., up to 95.45% accuracy for brain MR images).

Qualitatively, deep learning has enabled more automated and potentially more accurate analysis, reducing the burden on radiologists and improving the consistency of diagnoses.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Theoretical Understanding: The underlying theory of why deep learning models work so well, particularly regarding the optimal architecture and hyperparameters, is still not fully understood.
  • Data Scarcity: Insufficient labeled data remains a significant challenge in many medical imaging applications. Strategies like data augmentation, pretraining, and fine-tuning are helpful but their mechanisms require further investigation. Establishing public databases for molecular imaging data is also recommended.
  • Interpretability: The "black box" nature of deep learning models makes it difficult to understand how they arrive at their decisions. Establishing correlations between high-level features learned by DL models and actual clinical characteristics is crucial for increasing reliability.
  • Model Optimization: Further research is needed on optimizing model architectures and training strategies for specific tasks and datasets.
  • Unsupervised and Semi-supervised Learning: Given the cost and difficulty of manual annotation, developing more robust unsupervised and semi-supervised learning methods is essential.
  • Collaboration: Close collaboration between clinicians and engineers is vital to address these challenges and advance the application of deep learning in cancer molecular imaging.

Summary updated

Abstract

Molecular imaging enables the visualization and quantitative analysis of the alterations of biological procedures at molecular and/or cellular level, which is of great significance for early detection of cancer. In recent years, deep leaning has been widely used in medical imaging analysis, as it overcomes the limitations of visual assessment and traditional machine learning techniques by extracting hierarchical features with powerful representation capability. Research on cancer molecular images using deep learning techniques is also increasing dynamically. Hence, in this paper, we review the applications of deep learning in molecular imaging in terms of tumor lesion segmentation, tumor classification, and survival prediction. We also outline some future directions in which researchers may develop more powerful deep learning models for better performance in the applications in cancer molecular imaging.
BibTeX
@inproceedings{xue2017application,
  author = {Yong Xue and Shihui Chen and Jing Qin and Yong Liu and Bingsheng Huang and Hanwei Chen},
  title = {{Application of Deep Learning in Automated Analysis of Molecular Images in Cancer: A Survey}},
  booktitle = {Contrast},
  year = {2017},
  doi = {10.1155/2017/9512370},
}
Figure