PhyDCM: A Reproducible Open-Source Framework for AI-Assisted Brain Tumor Classification from Multi-Sequence MRI

PhyDCM is a reproducible, open-source framework that integrates a MedViT-based hybrid classification architecture with standardized DICOM processing and a modular desktop interface to achieve over 93% accuracy in AI-assisted brain tumor classification from multi-sequence MRI data.

The Gist

Imagine you are a doctor trying to diagnose a brain tumor. You have a massive stack of MRI scans (the "pictures" of the brain), and you need to figure out if there's a tumor, what kind it is (like a glioma, meningioma, or pituitary tumor), or if the brain is perfectly healthy.

Doing this manually is like trying to find a specific needle in a haystack while wearing thick gloves. It's slow, tiring, and easy to make mistakes when you have hundreds of patients.

This paper introduces PhyDCM, a new "smart assistant" designed to help doctors with this job. But instead of just being a magic black box that gives an answer, PhyDCM is built like a Lego set that anyone can take apart, study, and rebuild.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Black Box" Mystery

Currently, many AI tools for medicine are like sealed vending machines. You put an image in, and a result pops out. But you can't see inside, you can't change how it works, and if the machine breaks, you can't fix it because the code is hidden. This makes it hard for scientists to trust or improve them.

2. The Solution: The "Lego" Framework

The authors built PhyDCM as an open-source toolkit. Think of it as a high-end Lego set where every piece is labeled and accessible.

• The Brain (The Library): This is the "thinking" part. It does the heavy lifting of analyzing the images. It's written in a way that scientists can easily swap out pieces (like changing the brain's logic) without breaking the whole thing.
• The Face (The Desktop App): This is the screen doctors actually look at. It shows the MRI scans, lets you click through them, and displays the AI's diagnosis.
• The Magic: The best part is that the "Brain" and the "Face" are separate. You can use the Brain to process thousands of images automatically on a server, or use the Face to look at one image interactively. They don't depend on each other, making the system very flexible.

3. How It "Sees" (The MedViT Engine)

To understand the brain scans, PhyDCM uses a special AI engine called MedViT.

• The Old Way: Traditional AI looks at an image like a person looking at a painting through a tiny straw. It sees small details (edges, textures) but misses the big picture.
• The New Way (MedViT): This engine is like a smart detective. It uses two tools:
  1. A magnifying glass (Convolution): To look closely at small details like texture.
  2. A wide-angle lens (Transformer): To step back and see how different parts of the brain relate to each other.
     By combining these, it understands both the tiny details and the overall shape of a tumor, which helps it distinguish between different types of tumors more accurately.

4. The Training: Teaching the Assistant

The team taught this AI using a massive library of brain scans (over 6,000 images) from various sources.

• Standardizing the Input: Just like you wouldn't compare a photo taken in the dark with one in bright sunlight, the system first "normalizes" all the MRI scans. It adjusts the brightness and size so every image looks consistent before the AI studies it.
• The Test: They didn't just test it on the images it learned from (which would be cheating). They tested it on brand new, unseen datasets (like a final exam with new questions).
  • The Score: It got about 93% accuracy.
  • The Weakness: It sometimes got confused between a "Meningioma" (a specific type of tumor) and "No Tumor" because they can look very similar on a scan. But for the other types, it was nearly perfect.

5. Why This Matters

Most AI research papers just say, "We made a model that is 95% accurate." They don't give you the model, the code, or the interface.

PhyDCM is different because it gives you the whole kitchen, not just the recipe.

• Transparency: You can see exactly how it works.
• Reproducibility: If another scientist wants to test it, they can download the exact same code and get the same results.
• Future-Proof: Because it's built like a modular Lego set, if new types of scans (like CT or PET scans) become popular, scientists can just "plug in" a new module without rebuilding the whole system.

The Bottom Line

PhyDCM is a transparent, open-source "smart assistant" for brain tumor detection. It combines a powerful AI brain with a user-friendly interface, allowing doctors to see the scans and the diagnosis in one place. While it isn't ready to replace doctors in hospitals yet (it needs more testing and official approval), it provides a solid, trustworthy foundation for researchers to build the next generation of medical AI tools.

It turns the "black box" of AI into a glass house, where everyone can see how the magic happens.

Technical Summary

1. Problem Statement

The paper addresses two primary challenges in the current landscape of AI-driven medical imaging:

• Reproducibility and Accessibility: Many existing deep learning solutions for brain tumor classification are "closed" systems or monolithic codebases. They often lack transparency, making it difficult for researchers to modify, extend, or independently verify results. There is a disconnect between theoretical model performance and practical deployment.
• Workflow Integration: Most research focuses solely on the classification algorithm, neglecting the critical end-to-end pipeline required in clinical settings. This includes standardized DICOM processing, metadata handling, intensity normalization, and interactive visualization. Few open-source tools integrate raw DICOM data ingestion, preprocessing, AI inference, and structured reporting into a single, cohesive framework.

2. Methodology

The authors propose PhyDCM, a modular, open-source framework designed to bridge the gap between algorithmic research and clinical usability.

A. System Architecture

PhyDCM is structured as a dual-component system to ensure separation of concerns:

1. PhyDCM Python Library (Backend): A standalone computational module encapsulating all logic for image preprocessing, model management, and inference. It is designed to be reusable in command-line scripts or other environments without dependency on a graphical interface.
2. PhyDCM Desktop Application (Frontend): A PyQt5-based interactive application that delegates tasks to the library. It provides multi-planar visualization (axial, sagittal, coronal), user input management, and structured result display.

B. Core Components

• Model Architecture (MedViT Hybrid): The classification engine utilizes a MedViT (Medical Vision Transformer) architecture. This hybrid approach combines:
  • Convolutional Stems: For efficient local feature extraction (edges, textures).
  • Transformer Blocks: For modeling global spatial dependencies and long-range context, which pure CNNs often miss.
  • Classification Head: Global average pooling and fully connected layers for probability distribution.
• Preprocessing Pipeline:
  • DICOM Handling: Uses pydicom to parse metadata and pixel data.
  • Standardization: Intensity rescaling and resizing to 224x224 pixels.
  • Augmentation: Limited to training only (minor rotations, flips, zoom) to preserve diagnostic integrity.
• Automation: The system features an automatic model binding mechanism that dynamically scans directories for trained weights and label mappings, eliminating hard-coded paths and facilitating easy updates.

C. Datasets

• Training Data: A curated collection of ~6,650 MRI images from BRISC2025, FigShare, SARTAJ, and Br35H, covering four classes: Glioma, Meningioma, Pituitary Tumor, and No Tumor.
• Validation: The model was tested on the primary dataset and rigorously validated on two external, independent datasets (Nickparvar and Br35H) to assess generalizability across different acquisition protocols.

3. Key Contributions

1. Modular Open-Source Framework: Unlike monolithic tools, PhyDCM cleanly separates inference logic from the user interface, allowing independent development and extension.
2. End-to-End Pipeline: It integrates the entire diagnostic workflow: DICOM ingestion $\rightarrow$ Preprocessing $\rightarrow$ AI Inference $\rightarrow$ Structured Reporting (CSV/JSON).
3. Interactive Visualization: The desktop app includes synchronized multi-planar reconstruction (MPR), allowing users to navigate volumetric data and correlate visual findings with AI predictions in real-time.
4. Cross-Dataset Validation: The study emphasizes external validation on unseen datasets, a practice often missing in similar literature, to prove robustness against data distribution shifts.
5. Scalability: The architecture is designed to accommodate future modalities (CT, PET) and additional imaging types without rewriting core logic.

4. Results

The framework was evaluated on a total of over 3,000 test images across three datasets.

• Overall Performance: Achieved a combined classification accuracy of 93.33%.
• Dataset-Specific Accuracy:
  • BRISC2025: 92.30% (High accuracy on Glioma, Pituitary, and No Tumor; lower on Meningioma due to subtle morphological overlaps).
  • Nickparvar (External): 88.69% (Demonstrating strong generalization to external data).
  • Br35H (External): 100% accuracy on the "No Tumor" class tested.
• Confusion Analysis: The most frequent errors occurred between Meningioma and No Tumor categories, likely due to small or peripheral tumors mimicking normal tissue. Glioma and Pituitary tumors were classified with near-perfect accuracy.
• Comparison: While some existing studies report higher accuracy (e.g., 97%), they often use 3-class tasks or lack external validation. PhyDCM achieves competitive performance on a more complex 4-class task with rigorous cross-dataset testing.

5. Significance and Conclusion

The significance of PhyDCM lies in its holistic approach to medical AI. Rather than just proposing a new neural network, the authors provide a reproducible infrastructure that addresses the "last mile" problem in AI deployment:

• Transparency: By being open-source and modular, it allows the academic community to inspect, modify, and trust the diagnostic process.
• Usability: The integration of multi-planar visualization directly into the diagnostic tool reduces the cognitive load on specialists and bridges the gap between research and clinical practice.
• Future-Proofing: The design supports the addition of new modalities (CT, PET) and models, making it a sustainable foundation for future medical image analysis research.

Limitations: The authors note that while the framework is robust, current validation is limited to MRI. Expansion to CT and PET requires specific preprocessing and larger datasets. Furthermore, the system is currently a research/educational tool and requires further clinical validation and regulatory approval before hospital deployment.

In summary, PhyDCM represents a significant step toward standardized, transparent, and reproducible AI in medical imaging, prioritizing the integrity of the entire diagnostic pipeline over isolated algorithmic metrics.