Multi-grade brain tumor classification using deep CNN with extensive data augmentation

Problem and motivation

The WHO grades tumours of the central nervous system from I (benign) to IV (highly malignant), with grade IV glioblastomas carrying the worst prognosis. Grading normally requires histopathology — looking for necrosis, microvascular proliferation and vascular thrombosis — but those features are not always clear, and pathologists disagree about them. MRI offers a route to the same distinction without surgery, using contrasts that separate low-grade glioma from glioblastoma.

Computer-aided diagnosis systems built on MRI had three problems the paper sets out to fix. Most were binary — benign versus malignant — which leaves a radiologist without the information needed to choose a treatment. Public data is scarce for individual grades. And reported accuracy was not yet convincing for clinical use.

The framing matters: the system is proposed as a second opinion supporting a radiologist after clinical examination, not as a replacement for one.

How the method works

Three stages.

1. Tumour segmentation. InputCascadeCNN handles segmentation. Its distinguishing feature is two-pathway processing: a 7x7 receptive field stream extracting local features alongside a 13x13 stream for global context, with a final layer that implements the fully connected layer convolutionally — reported as roughly 40x faster than comparable models. Before the local pathway, N4ITK bias correction suppresses intensity extremes; before the global pathway, each channel is normalised by subtracting its mean and dividing by its standard deviation. Afterwards, connected-component labelling removes flat blobs that the bright skull margin would otherwise leave looking like tumour tissue.

2. Extensive data augmentation. Eight techniques split into two purposes:

• Geometric invariance — rotation (-90, -45, 45, 90 degrees), flips (top, bottom, left, right), skew (left, right, forward, backward), shear (10 degrees on X and Y).
• Noise invariance — Gaussian blur (sigma 0.25 to 2.0), sharpening (lightness 0.5 to 2.0), edge detection (alpha 0.25 to 1.0), emboss (strength 0.5 to 2.0).

Thirty parameter settings in total turn each MR image into 30 samples. Radiopaedia goes from 121 images to 3,630; the brain tumour dataset from 3,064 to 91,920.

3. Classification. VGG-19 is fine-tuned on the segmented, augmented tumour regions. The architectural argument for VGG-19 is specific to the domain: it uses 3x3 kernels with stride 1 throughout, where contemporaries used 11x11 or 7x7 kernels with strides of 4-5 in early layers. Large strides skip over exactly the fine texture that distinguishes one tumour grade from the next. Only the fully connected layer weights are updated; the 1,000-dimensional FC8 output feeds a SoftMax classifier over the four grades.

Accuracy before and after augmentation

Data was split 50% training / 25% validation / 25% testing, with each dataset evaluated both before and after augmentation so the augmentation effect is isolated.

On Radiopaedia, overall accuracy rises from 87.38% to 90.67%. Every grade improves, with the largest gain on grade I (90.03% to 95.54%) and the smallest on grade IV (85.50% to 86.71%).

On the brain tumour dataset (3,064 T1-weighted contrast-enhanced images from 233 patients, 512x512 at 0.49mm pixel size), sensitivity rises from 84.51% to 88.41% and specificity from 93.34% to 96.12%.

The residual confusion is clinically coherent rather than random. Grades III and IV are what get mixed up — before augmentation, 7.16% of grade III cases read as grade IV and 9.35% of grade IV as grade III. These are adjacent grades of the same glioma lineage, distinguished histologically by features that are subtle on MRI, and they remain the hardest pair after augmentation too.

Key contributions

• Presented what the authors describe as the first CAD system to classify brain tumours into all four WHO grades, rather than the binary benign/malignant split used in prior work.
• Addressed MRI data scarcity with a structured augmentation scheme covering both geometric and noise invariance, expanding each sample 30-fold and improving accuracy by 3.3 points.
• Argued and demonstrated that small 3x3 kernels at stride 1 matter for this task specifically, because tumour grade is signalled by fine texture that large-stride architectures skip.
• Reported before-and-after results on two datasets so the contribution of augmentation is separable from that of the architecture.

Datasets

• Radiopaedia — 121 MR images labelled by WHO grade — 36 grade I meningiomas, 32 grade II gliomas, 25 grade III gliomas, 28 grade IV glioblastomas. Expanded to 3,630 by augmentation.
• Brain tumour dataset (Cheng et al.) — 3,064 T1-weighted contrast-enhanced images from 233 patients at 512x512, 0.49mm pixels, 6mm slices — 708 meningioma, 1,426 glioma, 930 pituitary. Expanded to 91,920.

Results

Per-grade accuracy on Radiopaedia before and after augmentation, and overall sensitivity/specificity on the brain tumour dataset.

Limitations and open questions

• Augmentation multiplies samples but not information. With 121 source images on Radiopaedia, train and test augmentations derive from the same small pool of patients, so the reported gain reflects invariance to transformation rather than exposure to genuinely new cases.
• Grades III and IV remain the dominant confusion — precisely the boundary that matters most clinically, since it separates anaplastic glioma from glioblastoma.
• The two datasets are labelled differently: Radiopaedia by WHO grade, the brain tumour dataset by tumour type (meningioma, glioma, pituitary), so the two evaluations are not directly comparable and only the first tests the four-grade claim.
• Only the fully connected layers are fine-tuned, leaving the convolutional features shaped entirely by natural images rather than by MRI.
• No prospective or cross-institution validation is reported, so performance on scanners and protocols outside these datasets is unknown.

Related work on this site

• Efficient image recognition and retrieval on IoT-assisted energy-constrained platforms from big data repositories
• Integrating salient colors with rotational invariant texture features for image representation in retrieval systems
• Densely deformable efficient salient object detection network