BellCNN: CNN for Alzheimer’s Detection

• The paper introduces BellCNN's innovative bell-shaped filter progression for binary classification of Alzheimer’s disease in MRI images.
• BellCNN is specifically engineered for the OASIS MRI dataset using dichotomized Clinical Dementia Rating labels, contrasting with transfer-learning approaches.
• Using customized preprocessing and dropout regularization, BellCNN achieved over 95% training accuracy and promising results on unseen test images.

Searching arXiv for BellCNN and closely related naming variants to ground the article. BellCNN is a custom convolutional neural network introduced for automatic detection of Alzheimer’s disease from MRI in the paper "Detection of Alzheimers Disease from MRI using Convolutional Neural Networks, Exploring Transfer Learning And BellCNN" (Awate, 2019). In that work, BellCNN is a bespoke 2D CNN for binary classification of MRI images into Alzheimer’s disease versus control, built specifically for the OASIS cross-sectional MRI dataset rather than adopted as a standard architecture from the broader literature. The model is framed as an alternative to transfer-learning baselines based on Inception and MobileNet, with the stated motivation that a network built "from the ground up" for MRI may be less affected by "foreign input spaces" than pretrained natural-image models (Awate, 2019).

1. Definition and problem setting

BellCNN is defined in the paper as a custom-made CNN for MRI-based Alzheimer’s disease detection, with the learning task reduced to binary classification using Clinical Dementia Rating labels (Awate, 2019). The dataset is the OASIS cross-sectional MRI collection, described as containing 416 subjects, all right-handed, with standardized T1-weighted MRI images. According to the paper, 100 subjects have mild to moderate dementia, and the cohort includes both younger non-demented adults and 98 elderly non-demented subjects over age 60.

The labeling rule used for BellCNN is explicit. The paper maps CDR = 0 to no dementia / control and CDR > 0 to dementia / Alzheimer’s disease. BellCNN therefore does not perform multi-class severity prediction. It is a binary detector that collapses the original dementia spectrum into control versus dementia.

The paper also situates BellCNN within a broader diagnostic motivation. MRI reveals atrophied regions associated with Alzheimer’s disease, but those regions differ across individuals, which the authors identify as a source of diagnostic difficulty. BellCNN is accordingly presented as a pre-flagging model that could assist further clinical assessment rather than as a replacement for clinical diagnosis. A plausible implication is that the intended use is triage or screening support rather than full severity staging, because the learning target is binary and the evaluation does not address clinical-grade deployment.

2. Architectural design and the origin of the name

BellCNN is not described as a standard named network family such as ResNet or VGG. Its name comes from the pattern of convolutional filter counts, which increase and then decrease in a bell-shaped progression:

$$32 \rightarrow 64 \rightarrow 128 \rightarrow 64 \rightarrow 32$$

The paper explicitly states that the architecture "takes inspiration from the shape of the Bell Curve in Gaussian distribution" (Awate, 2019).

At a high level, the network contains an input layer, five convolutional layers, max pooling after each convolution, a fully connected layer of 1024 nodes, dropout, a SoftMax stage, and a final output layer of 2 nodes. ReLU is described as the primary activation in the hidden pipeline, and Adam is the optimizer.

The layer sequence reported in the paper is:

1. Convolutional layer with 32 filters
2. Max pooling
3. Convolutional layer with 64 filters
4. Max pooling
5. Convolutional layer with 128 filters
6. Max pooling
7. Convolutional layer with 64 filters
8. Max pooling
9. Convolutional layer with 32 filters 10. Max pooling
10. Flatten / fully connected layer with 1024 nodes
11. Dropout rate = 0.8
12. SoftMax
13. Final output layer with 2 nodes
14. Adam optimizer
15. Regression / training

The pseudo-algorithm additionally states "Set stride to 5" and "Set activation to RELU" before the convolutional sequence (Awate, 2019). The architecture is therefore defined by a clear filter-width schedule and stage order, but not by a complete low-level specification.

Several implementation details are explicitly absent or ambiguous. The paper does not give the convolution kernel size, pooling window size, pooling stride, padding type, input image dimensions, or a precise statement of whether the declared stride of 5 applies uniformly to all convolutional layers. The meaning of "dropout rate of 80% (0.8)" is also internally inconsistent, because the accompanying text explains that each node is active 80% of the time and off 20% of the time. This suggests that the intended parameter is keep probability 0.8 rather than modern drop rate 0.8.

3. Data representation and preprocessing

The BellCNN paper reports limited but concrete preprocessing steps (Awate, 2019). The authors performed data wrangling, created randomized arrays of image data, saved those arrays as NumPy arrays, and converted labels to one-hot encoding. They also state that they used processed images provided within OASIS and performed pre-labeling based on CDR, with some data preserved as a test set.

The paper does not specify image resolution after loading, any resizing procedure, intensity normalization, skull stripping, registration details beyond use of processed OASIS images, or whether the network uses full MRI volumes, slices, or another derived representation. The description consistently refers to "images" and presents BellCNN as a standard 2D CNN pipeline. This suggests 2D image inputs rather than full 3D MRI volumes, but that remains an inference rather than a formally stated design choice.

The train/validation/test protocol is also only partially specified. The paper says that some of the data is preserved as a test dataset. For the transfer-learning experiments, the final test set size is explicitly reported as $N = 98$, but no equally complete split description is given for BellCNN itself. The reported BellCNN training regime is "more than 100 epochs," with each epoch consisting of 7 batches, and performance becoming consistently good from around step 500 (Awate, 2019).

A plausible implication is that BellCNN’s reproducibility is constrained less by the macro-architecture than by missing preprocessing and split details. The high-level pipeline is identifiable, but exact reconstruction from the paper alone is not possible.

4. Optimization, loss, and reported empirical behavior

The training details that are explicit are narrow but definite. BellCNN is trained with the Adam optimizer, ReLU hidden activations, SoftMax output, categorical cross-entropy loss, and dropout regularization (Awate, 2019). The software stack includes TensorFlow, TensorBoard, and the Anaconda / Python ecosystem, and the paper refers to freezing trained graphs as protobuf $.pb$ files.

The paper reports training-curve behavior rather than a full benchmark table for BellCNN. Specifically, it states that:

• BellCNN was trained for more than 100 epochs
• Each epoch consisted of 7 batches
• From about step 500, results became consistently good
• Initial loss was "as high as 60%"
• Loss reduced to less than 10% after about step 500
• Accuracy increased to more than 95% from about step 500 onward

These are training metrics, not comprehensive held-out evaluation metrics. The paper does not report full test accuracy on the entire test split, nor sensitivity, specificity, precision, recall, AUC, or a complete confusion matrix for BellCNN (Awate, 2019).

The only explicit held-out qualitative test example for BellCNN is a set of 9 randomly selected test images that the model had never seen before, consisting of 3 Alzheimer’s disease images and 6 control images. The paper reports that BellCNN correctly predicted all 9. This is consistent with the authors’ claim that the model can distinguish dementia from control on unseen images, but it is not a substitute for a full test-set evaluation.

5. Relation to transfer learning and evaluation caveats

The same paper evaluates transfer-learning pipelines based on Inception and MobileNet using TensorFlow retraining (Awate, 2019). Their reported final test accuracies are:

Model	Reported final test accuracy	Test size
Inception	85.7%	$N = 98$
MobileNet	81.6%	$N = 98$

The paper also reports example inference times of about 4.0 s/image for Inception and about 1.45 s/image for MobileNet. BellCNN is then argued to be preferable because it is task-specific and less likely to inherit biases from pretrained models operating on unrelated datasets.

That conclusion, however, is methodologically limited by the asymmetry of the evaluation. Transfer learning receives a clearly reported final test accuracy on $N = 98$, whereas BellCNN is characterized mainly by training curves and perfect classification on 9 selected unseen test images. The paper therefore implies that BellCNN is competitive with or superior to the transfer-learning baselines, but it does not demonstrate that claim under an equivalent test protocol.

This asymmetry is the principal controversy surrounding BellCNN in the paper’s presentation. The architecture itself is straightforward, and the training behavior is reported as favorable, but the evidence is not standardized across methods. For that reason, BellCNN is best interpreted as a promising research prototype or pre-flagging tool rather than a clinically validated diagnostic system. The paper’s own tone is optimistic about future deployment, but the reported evidence does not establish deployment readiness.

6. Limitations, ambiguities, and scope

BellCNN’s main strengths within the paper are its task-specific construction, regularization through dropout, and the reported combination of low training loss and high training accuracy after approximately 500 steps (Awate, 2019). The authors also explicitly credit dropout regularization and efficient data wrangling for helping obtain a "more generalized model."

Its limitations are equally explicit. The architecture description omits essential hyperparameters such as kernel sizes, padding, pooling specifications, and batch size. The input representation is not described with sufficient precision to determine whether the model consumes slices, selected views, or another 2D derivation of MRI data. The class balance after preprocessing is not reported, and the binary label construction collapses dementia severity into a single positive class. The evaluation methodology for BellCNN is weaker than for the transfer-learning baselines, because no full test metric over the complete test split is given.

These omissions matter for interpretation. The paper demonstrates that BellCNN can be trained in TensorFlow, can achieve more than 95% training accuracy after about step 500, and can correctly classify 9 selected unseen test images. It does not demonstrate calibrated clinical performance, robustness across sites, or comparative superiority under a unified experimental protocol. This suggests that BellCNN should be understood as an exploratory CNN architecture for MRI-based Alzheimer’s disease detection rather than as an established benchmark model.

7. Terminological disambiguation

"BellCNN" is a genuine model name in the Alzheimer’s MRI paper (Awate, 2019), but it is also easily confused with several similarly named arXiv models that are unrelated in task and architecture.

LoadCNN is not BellCNN. "LoadCNN: A Low Training Cost Deep Learning Model for Day-Ahead Individual Residential Load Forecasting" proposes a CNN-based model for next-day forecasting of individual residential electricity demand and explicitly states that the paper does not present a model called BellCNN (Huang et al., 2019).

BCN is not BellCNN. "Broadcasting Convolutional Network for Visual Relational Reasoning" introduces the Broadcasting Convolutional Network as a module for global feature extraction, spatial encoding, and broadcast redistribution in visual relational reasoning (Chang et al., 2017).

BCNN is not BellCNN. "BCNN: Binary Complex Neural Network" proposes a binary complex extension of binarized neural networks and explicitly notes that the paper does not mention a model called BellCNN (Li et al., 2021).

B-CNN is not BellCNN. "B-CNN: Branch Convolutional Neural Network for Hierarchical Classification" introduces a hierarchy-aware CNN with multiple branch outputs and similarly clarifies that "BellCNN" is not the name used in that paper (Zhu et al., 2017).

Within the arXiv record considered here, BellCNN therefore refers specifically to the custom Alzheimer’s MRI classifier from (Awate, 2019). Its defining characteristics are the binary CDR-based labeling scheme, the bell-shaped filter sequence $$32 \rightarrow 64 \rightarrow 128 \rightarrow 64 \rightarrow 32$$, the 1024-unit fully connected layer, dropout specified as 0.8 in the paper’s terminology, SoftMax output over 2 nodes, and evaluation centered on training curves plus limited held-out image examples.