By Harshitha Govindaraju · Hassan Lab · Rutgers ECE
A convolutional neural network trained on micrographs of human leukocytes and spherical particles. Drop a fluorescent image — the model slices it into roughly 100 px tiles, regresses a per-tile count, and sums them.
Estimated cell count
0
Ceiling of summed per-tile predictions.
Method
The architecture is a sequential convolutional network with four
progressively deeper blocks. Each block holds two or three 3×3
convolutions, max-pooled and dropped at 0.3. The dense
head regresses a single non-negative scalar — the predicted cell
count for that 128×128 tile.
A ZeroPadding2D(40, 40) after the first convolution
expands the receptive field before the deeper layers compress it.
The full image is split into a uniform grid of approximately
100 px tiles, each resized to 128×128×3, batched through the
network. The total cell count is the ceiling of the summed
predictions.
Training: MSE loss under Adam at lr = 1e-3,
batch size 16, on fluorescent micrographs of human leukocytes and
spherical particles. Architecture, dataset curation, and training
pipeline by Harshitha Govindaraju (M.S. thesis,
Rutgers, 2021).
Layer Output input 128 × 128 × 3 conv2d 32f 3×3 126 × 126 × 32 zeropad +40 206 × 206 × 32 conv2d 32f 3×3 204 × 204 × 32 conv2d 32f 3×3 202 × 202 × 32 maxpool 2×2 101 × 101 × 32 dropout 0.3 ───────────────────────────────────── conv2d 64f 3×3 99 × 99 × 64 conv2d 64f 3×3 97 × 97 × 64 conv2d 64f 3×3 95 × 95 × 64 maxpool 2×2 47 × 47 × 64 dropout 0.3 ───────────────────────────────────── conv2d 128f 3×3 45 × 45 × 128 conv2d 128f 3×3 43 × 43 × 128 conv2d 128f 3×3 41 × 41 × 128 maxpool 2×2 20 × 20 × 128 dropout 0.3 ───────────────────────────────────── conv2d 128f 3×3 18 × 18 × 128 conv2d 128f 3×3 16 × 16 × 128 conv2d 128f 3×3 14 × 14 × 128 maxpool 2×2 7 × 7 × 128 dropout 0.3 ───────────────────────────────────── flatten 6272 dense + LReLU + BN 1024 dense + LReLU + BN 512 dense + ReLU 1 ───────────────────────────────────── total parameters 7,878,529
Best results
The model was trained on fluorescent images of human leukocytes and spherical particles at resolutions above 1024² px. This is a research demonstration — not a clinical instrument.
Research
The convolutional approach to leukocyte counting on smartphone-coupled microfluidic biosensors was developed in this body of work. The IEEE Access 2022 paper is the primary publication for this counter.
Author
Lead author · Architecture · Training · Web application
M.S., Biomedical Engineering · Rutgers University
Designed the convolutional architecture, curated and labeled the
training dataset, ran the training pipeline, and built the original
web application. The cell_count_v2 model and the 2021
Rutgers M.S. thesis on which this site is based are her work.
Acknowledgments
Image acquisition support · Ph.D., Electrical & Computer Engineering, Rutgers
Faculty advisor · Assistant Professor, Hassan Lab, Rutgers ECE