Classification of tumor from computed tomography images: A brain-inspired multisource transfer learning under probability distribution adaptation

Yu Liu, Enming CuiFrontiers in Human Neuroscience 2022 · DOI

Abstract

Preoperative diagnosis of gastric cancer (GC) and primary gastric lymphoma (PGL) is challenging and has important clinical significance. Inspired by the inductive reasoning learning of the human brain, transfer learning can improve diagnosis performance of target task by utilizing the knowledge learned from the other domains (source domain). However, most studies focus on single-source transfer learning and may lead to model performance degradation when a large domain shift exists between the single-source domain and target domain. By simulating the multi-modal information learning and transfer mechanism of human brain, this study designed a multisource transfer learning feature extraction and classification framework, which can enhance the prediction performance of the target model by using multisource medical data (domain). First, this manuscript designs a feature extraction network that takes the maximum mean difference based on the Wasserstein distance as an adaptive measure of probability distribution and extracts the domain-specific invariant representations between source and target domain data. Then, aiming at the random generation of parameters bringing uncertainties to prediction accuracy and generalization ability of extreme learning machine network, the 1-norm regularization is used to implement sparse constraints of the output weight matrix and improve the robustness of the model. Finally, some experiments are carried out on the data of two medical centers. The experimental results show that the area under curves (AUCs) of the method are 0.958 and 0.929 in the two validation cohorts, respectively. The method in this manuscript can provide doctors with a better diagnostic reference, which has certain practical significance.

Introduction: The Diagnostic Challenge

Distinguishing between Gastric Cancer (GC) and Primary Gastric Lymphoma (PGL) is a high-stakes challenge for radiologists. While they look similar on a CT scan, their treatments are polar opposites: GC usually requires surgery, whereas PGL is best treated with chemotherapy or radiotherapy. An incorrect preoperative diagnosis can lead to unnecessary or inappropriate invasive procedures.

Deep learning offers a solution, but it faces a classic medical AI hurdle: data scarcity. Standard transfer learning—taking a model trained on one dataset (like ImageNet) and fine-tuning it for another—often fails when the "source" and "target" domains are too different, a phenomenon known as "negative transfer." This paper seeks to bridge that gap by mimicking the human brain’s ability to integrate diverse, multi-modal information to make decisions.

Methodology: The AMSF-L1ELM Framework

The researchers proposed the Adaptation Multiple Spaces Feature and L1-norm Regularization Extreme Learning Machine (AMSF-L1ELM). This brain-inspired architecture consists of three main stages designed to maximize knowledge transfer from multiple sources.

Architecture of the AMSF-L1ELM framework, showing multisource feature extraction and the ELM classification layer.
  1. Multisource Feature Extraction: The model utilizes three distinct sources: gastric whole-slide images (WSI), lung WSIs, and arterial-phase CT scans. A ResNet50 backbone maps these into a common feature space.
  2. Adaptive Distribution Alignment: The model uses Maximum Mean Discrepancy (MMD) based on Wasserstein distance to calculate a probability adaptation factor ($\mu$). This factor dynamically decides whether to focus on aligning the global data distribution (marginal) or the class-specific distribution (conditional) for each source-target pair.
  3. Robust Classification: Features are processed by an Extreme Learning Machine (ELM). To address the instability of random weight initialization, the authors added $L_1$-norm regularization, enforcing sparsity and improving model robustness.

Data Sources and Domain Adaptation

The study utilized a diverse array of datasets to train and validate the multisource approach, including gastric cancer WSIs, lung WSIs from TCGA, and gastric artery phase CT images. The target domain consisted of gastric venous phase CT images from two different medical centers.

Visualization of distribution shifts across different source domains and the target domain, highlighting the need for adaptive alignment.

A key insight from the study is the "Power of Diversity." Including lung tissue data (S2) improved the classification of gastric tumors, suggesting the model learned fundamental medical features that generalize across organs. The dynamic weighting ($\mu$) acknowledges that not all source data is equally relevant; for instance, arterial CTs are more similar to target venous CTs than WSIs are.

Results and Clinical Impact

The AMSF-L1ELM model significantly outperformed traditional clinical models and state-of-the-art domain adaptation methods:

  • Accuracy: It achieved an AUC of 0.958 on the internal validation cohort and 0.929 on the external validation cohort.
  • Comparison: It outperformed established methods like DAAN (AUC 0.859) and DSAN (AUC 0.783).
  • Clinical Relevance: With high sensitivity (0.980) and specificity (0.920), the model could serve as a reliable second opinion for doctors, potentially reducing misdiagnosis rates in clinical practice.

Conclusion

While the results are impressive, the authors note that a significant distribution gap remains between some sources and the target domain. Future work will focus on developing more sophisticated distance metrics to bridge these domain shifts and conducting broader multi-center trials to ensure cross-platform reliability.

Summary updated

Abstract

Preoperative diagnosis of gastric cancer and primary gastric lymphoma is challenging and has important clinical significance. Inspired by the inductive reasoning learning of the human brain, transfer learning can improve diagnosis performance of target task by utilizing the knowledge learned from the other domains (source domain). However, most studies focus on single-source transfer learning and may lead to model performance degradation when a large domain shift exists between the single-source domain and target domain. By simulating the multi-modal information learning and transfer mechanism of human brain, this study designed a multisource transfer learning feature extraction and classification framework, which can enhance the prediction performance of the target model by using multisource medical data (domain). First, this manuscript designs a feature extraction network that takes the maximum mean difference based on the Wasserstein distance as an adaptive measure of probability distribution and extracts the domain-specific invariant representations between source and target domain data. Then, aiming at the random generation of parameters bringing uncertainties to prediction accuracy and generalization ability of extreme learning machine network, the 1-norm regularization is used to implement sparse constraints of the output weight matrix and improve the robustness of the model. Finally, some experiments are carried out on the data of two medical centers. The experimental results show that the area under curves (AUCs) of the method are 0.958 and 0.929 in the two validation cohorts, respectively. The method in this manuscript can provide doctors with a better diagnostic reference, which has certain practical significance.
BibTeX
@inproceedings{liu2022classification,
  author = {Yu Liu and Enming Cui},
  title = {{Classification of tumor from computed tomography images: A brain-inspired multisource transfer learning under probability distribution adaptation}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2022},
  doi = {10.3389/fnhum.2022.1040536},
}
Figure