You can deploy this model and start earning money today!

rarfileexe/
xpathology-brain-specialist

Name: xpathology-brain-specialist
Rating: 2.5 (7 reviews)

computational-neurology

masoudnickparvar/brain-tumor-mri-dataset

apache-2.0

Views

No views yet

XPathology — Brain Tumor MRI Specialist (EfficientNetB0)

Part of the X-Pathology AI-Assisted Medical Screening Platform
_{Developed by Muhammad Hassan · AgenticEra Systems}

Overview

XPathology Brain Tumor MRI Specialist is a fine-tuned EfficientNetB0 CNN trained to classify brain MRI scans into 4 diagnostic categories: glioma, meningioma, no tumor, and pituitary tumor. It is the second specialist model in the X-Pathology platform, which combines CNN classification with Grad-CAM explainability and Gemini-powered dual clinical/patient reporting.

This model was trained and evaluated entirely on the masoudnickparvar/brain-tumor-mri-dataset, using its pre-defined Training/ and Testing/ splits — the standard academic partition for this benchmark dataset.

You can see the whole Training Notebook of this model: https://github.com/Muhammad-Hassan12/xpathology-brain-specialist

Disclaimer: This model is intended for research and educational use only. It is not FDA-approved or CE-marked diagnostic software. All outputs must be verified by a licensed radiologist or neurosurgeon before any clinical decision-making.

Model Details

Property	Value
Model name	XPathology_BrainSpecialist_B0_v2
Architecture	EfficientNetB0 (ImageNet pretrained)
Framework	TensorFlow / Keras
Input size	224 × 224 × 3 (RGB, pixel values 0–255)
Output	4-class softmax
Training split	masoudnickparvar Training/ (5,600 images, 80/20 internal split)
Holdout split	masoudnickparvar Testing/ (1,600 images, pre-defined by dataset authors)
Temperature (T)	0.7867 (post-hoc calibration applied)
Mixed precision	Yes (float16 compute, float32 weights)
Training hardware	Kaggle T4 (single GPU)

Tumor Classes

Index	Class	Description
0	glioma	Malignant tumor arising from glial cells — most aggressive class
1	meningioma	Typically benign tumor of the meninges
2	notumor	No tumor detected — healthy brain tissue
3	pituitary	Benign adenoma of the pituitary gland

Performance

Internal Validation (20% split of Training/)

Metric	Value
Accuracy	94.20%
AUC	0.9942
Top-2 Accuracy	98.93%
Val Loss	0.5124

Holdout Validation (Testing/ — 1,600 images, never seen during training)

Metric	Value
Accuracy	90.06%
AUC	0.9786
Macro F1	0.8982
Domain shift gap	4.14%

Per-Class Performance (Holdout)

Class	Precision	Recall	F1	Support
glioma	0.9228	0.7775	0.8440	400
meningioma	0.8942	0.8450	0.8689	400
notumor	0.8980	0.9900	0.9417	400
pituitary	0.8919	0.9900	0.9384	400

Note on glioma recall (77.75%): Glioma is the most morphologically heterogeneous class. The glioma↔meningioma confusion (39 cases) is consistent with published benchmarks on this dataset — both tumor types can present with overlapping MRI signal characteristics at 2D slice level. This was confirmed across four systematic training runs (varying architecture, unfreeze percentage, and loss function), all showing glioma recall in the 77–80% range, confirming a dataset-level ceiling rather than a training failure.

Training Configuration

Architecture

Input (224×224×3)
  └── EfficientNetB0 backbone (ImageNet weights)
        └── GlobalAveragePooling2D
              └── BatchNormalization
                    └── Dropout(0.5)
                          └── Dense(4, softmax, dtype=float32)

Two-Phase Training

Phase 1 — Warm-up (10 epochs)

Backbone fully frozen
Adam optimizer, LR = 1e-3
CategoricalCrossentropy, label smoothing = 0.1
EarlyStopping (patience=5)

Phase 2 — Fine-tuning (30 epochs, early stopping)

Top 20% of backbone layers unfrozen
All BatchNorm layers remain frozen throughout
Adam optimizer, LR = 1e-5, gradient clipping (clipnorm=1.0)
EarlyStopping (patience=5), ReduceLROnPlateau (factor=0.3, patience=3)
Best checkpoint saved on val_loss

MRI-Specific Augmentation (CPU-side via tf.data)

RandomFlip (horizontal only)      — brains are left-right symmetric
RandomRotation (±15°)             — small rotation, MRI has fixed orientation
RandomBrightness (±0.15)          — scanner intensity variation
RandomContrast (0.85–1.15)        — scanner contrast variation
RandomCrop (95–100%)              — simulates scan field-of-view variation
NO vertical flip                  — anatomically invalid for brain MRI
NO hue/saturation augmentation    — MRI is grayscale-to-RGB, colour is meaningless

Confidence Calibration

Post-training temperature scaling was applied on the holdout set to produce calibrated probabilities:

Optimal temperature T = 0.7867
NLL before calibration: 0.3479
NLL after  calibration: 0.3335

T < 1 indicates the raw softmax outputs were more diffuse than the ground truth distribution — consistent with label smoothing during training. Apply at inference: probs = softmax(logits / T).

Usage

Installation

pip install tensorflow>=2.12 numpy opencv-python

Load and Infer

python

import numpy as np
import tensorflow as tf
import json
import cv2

model   = tf.keras.models.load_model('xpathology_brain_specialist_b0.keras')
T       = np.load('brain_temperature_value.npy')[0]
classes = json.load(open('brain_class_names.json'))

logit_model = tf.keras.Model(
    inputs  = model.input,
    outputs = model.get_layer('uncertainty_dropout').output
)
final_dense   = model.get_layer('brain_specialist_output')
weights, bias = final_dense.get_weights()

def predict(image_array):
    """
    Args:
        image_array: np.ndarray shape (224, 224, 3), float32, range [0, 255]
    Returns:
        predicted_class (str), confidence (float), all_probs (dict)
    """
    img      = tf.expand_dims(image_array, axis=0)
    features = logit_model(img, training=False).numpy()
    logits   = features @ weights + bias
    scaled   = logits / T
    exp_l    = np.exp(scaled - scaled.max())
    probs    = (exp_l / exp_l.sum()).flatten()
    idx      = int(np.argmax(probs))
    return classes[idx], float(probs[idx]), dict(zip(classes, probs.tolist()))


# Example
img = cv2.imread('brain_mri.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (224, 224)).astype(np.float32)

label, conf, probs = predict(img)
print(f"Prediction : {label}")
print(f"Confidence : {conf:.4f}")
print(f"All probs  : {probs}")

Input Requirements

Property	Requirement
Modality	T1 or T2-weighted brain MRI
Orientation	Axial plane preferred
Resolution	Any (resized to 224×224)
Color space	RGB (convert grayscale to 3-channel first)
Pixel range	0–255 — do NOT normalize to [0,1]
File formats	JPEG, PNG

Important: EfficientNetB0 performs its own internal normalization. Do not apply rescale=1./255 before passing the image to the model.

Repository Structure

xpathology-brain-specialist/
├── xpathology_brain_specialist_b0.keras   # Trained model weights
├── brain_temperature_value.npy            # Calibration temperature T = 0.7867
├── brain_class_names.json                 # Ordered class labels
├── brain_training_summary.json            # Full training metadata
├── brain_training_log.csv                 # Epoch-by-epoch metrics (40 epochs)
└── README.md

Limitations

2D single-slice only: This model classifies individual 2D MRI images. Clinical diagnosis uses full 3D volumetric context across multiple slices and modalities (T1, T2, FLAIR, T1ce). Slice-level classification is a known simplification.
Glioma recall ceiling ~78%: Documented across four systematic training runs. The glioma↔meningioma overlap is inherent to 2D T1-weighted MRI without multi-modal or 3D context.
Single dataset source: Training and Testing splits originate from the same dataset. Cross-institutional validation has not been performed.
Axial plane orientation: Trained on axial-plane images. Coronal or sagittal slices may produce unreliable predictions.
Not for clinical use: This model has not undergone regulatory review and must not be used as the sole basis for any diagnostic or treatment decision.

X-Pathology Platform

This is the second specialist in X-Pathology, a multi-modal AI medical screening research platform.

Deployed specialists:

Specialist	Architecture	Val accuracy	Holdout accuracy
Colon Histopathology	EfficientNetB1, 9 classes	99.1%	92.7% (external dataset)
Brain MRI (this model)	EfficientNetB0, 4 classes	94.2%	90.1%

Pipeline:

User uploads brain MRI scan
EfficientNetB0 classifies tumor type
Grad-CAM generates attention heatmap
Gemini 2.5 Flash writes dual report — clinical + plain-English patient summary

Live demo: x-pathology.vercel.app GitHub: github.com/Muhammad-Hassan12

Training Reproducibility

Python          3.12
TensorFlow      2.x
Keras           3.x
Hardware        Kaggle T4 (single GPU — no MirroredStrategy)
Mixed precision float16
Batch size      32

License

Released under the Apache 2.0 License. The training dataset is provided by Masoud Nickparvar on Kaggle. Please credit the dataset source in published work.

Citation

bibtex

@misc{hassan2026xpathologybrain,
  author    = {Muhammad Hassan},
  title     = {XPathology Brain Tumor MRI Specialist: EfficientNetB0 for 4-Class Brain MRI Classification},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/rarfileexe/xpathology-brain-specialist}
}

Built by Muhammad Hassan · AgenticEra Systems
_{BSCS Student · Certified GenAI Developer · Computational Pathology & Radiology Research}