Abstract
Deep learning models trained on imbalanced data often perform poorly as they tend to be dominated by majority classes, potentially ignoring minority classes with limited data. This paper presents the first application of deep learning-based imbalanced data classification to cellular macromolecular complexes captured by Cryo-electron tomography (Cryo-ET). We employ a suite of strategies, including data sampling, bagging, boosting, and Genetic Programming. Notably, we propose a multi-path CNN model inspired by the Inception 3D network, integrated with focal loss and mixup data augmentation. In this architecture, each path is optimized for a specific data type, allowing the network to learn optimal path combinations. Extensive experiments demonstrate the flexibility of our approach across varying class counts. This work establishes a new path for cell classification in computational biology by effectively addressing extreme data imbalance in internal tissue imaging.
Introduction
Cellular Electron Cryotomography (Cryo-ET) is a transformative imaging technique that captures high-resolution, 3D visualizations of cellular structures in their near-native states. It is essential for discovering and analyzing macromolecular complexes within single cells. However, a significant hurdle in this field is the extreme biological variation in protein concentrations, which can span seven orders of magnitude. This leads to Cryo-ET datasets that are inherently and severely imbalanced.
Standard deep learning classifiers typically fail in such scenarios, favoring abundant proteins (majority classes) while neglecting rare ones (minority classes). While the computer vision community has developed various imbalanced learning techniques, their application to the internal tissue classification of cell macromolecules has been limited. This paper introduces a novel deep learning strategy that combines a custom "multi-path" Convolutional Neural Network (CNN) with mixup data augmentation and focal loss to overcome these challenges.
Methodology
The proposed approach addresses data imbalance through interventions at the data, loss, and architectural levels.
Overall Workflow
The classification pipeline begins with the CECT dataset, followed by data augmentation via mixup. The augmented data passes through initial convolutional and activation layers before entering the core Multi-Path Model. The output is then processed through activation scaling, pooling, and fully connected layers, with focal loss applied during training to guide the final softmax classification.
Multi-Path CNN Architecture
Inspired by the Inception 3D network, the core of our strategy is a CNN with multiple parallel paths. The number of paths is dynamically adjusted to match the number of classes in the dataset. This design effectively decomposes the multi-class problem into several binary classification tasks. Each path is trained to become an "expert" at identifying a specific macromolecule class. The features extracted from these parallel paths are concatenated and passed through subsequent layers for the final decision.
Mixup and Focal Loss
To further enhance performance:
- Mixup Data Augmentation: Virtual training examples are created by linearly interpolating pairs of input images and their labels, expanding the training distribution and preventing overfitting on minority classes.
- Focal Loss: We replace standard cross-entropy with an $\alpha$-balanced focal loss. This reshapes the loss function to down-weight easy, well-classified examples from majority classes, forcing the network to focus on hard, misclassified examples from minority classes.
Datasets and Experimental Setup
The model was evaluated on a real-world Cryo-ET dataset extracted from a rat neuron tomogram. The dataset consists of 3D subtomograms (binned to $40^3$ voxels) representing four types of macromolecules with significant class imbalance:
- Double-capped proteasome (proteasome_d): 1043 samples (63.5%)
- Single-capped proteasome (proteasome_s): 386 samples (23.5%)
- TRiC (TCP-1 ring complex): 125 samples (7.6%)
- Ribosome: 80 samples (4.9%)
The data was split into 1307 training and 327 testing samples. Performance was measured using Macro F1-Score and Macro G-mean to ensure a balanced evaluation across all classes, as standard accuracy can be misleading in imbalanced contexts.
Results and Discussion
The proposed Multi-path CNN + mixup + focal loss strategy achieved superior results, yielding a Macro F1 score of 73.6 and a Macro G-mean of 84.7. This performance significantly outperformed traditional baselines such as Granular Support Vector Machine with Repetitive Undersampling (GSVM-RU), which achieved a Macro F1 of 69.5 and a Macro G-mean of 80.3.
Key insights from the study include:
- Architectural Specialization: Dedicating specific CNN paths to each class prevents minority features from being overwhelmed by majority classes during training, proving more effective than generic resampling methods like SMOTE.
- Synergistic Effects: The combination of data-level (mixup), loss-level (focal loss), and architectural (multi-path) interventions was crucial for achieving peak performance.
- Flexibility: Ablation studies confirmed the model's adaptability, as it consistently outperformed baselines when adjusted for different numbers of classes.
Conclusion
This research demonstrates that a specialized deep learning architecture combined with targeted data and loss strategies can effectively classify macromolecular complexes in highly imbalanced Cryo-ET datasets. While the multi-path approach scales linearly with the number of classes—posing potential computational challenges for extremely large class sets—it represents a significant step forward in computational biology. Future work will focus on improving the scalability of this framework and applying it to broader biological imaging challenges.
Summary updated