Deep Learning-Based Blood Cell Detection in Microscopic Images for Enhanced Disease Recognition with RetinaNet

A complete blood count — how many red cells, white cells, and platelets are in a sample — drives a large share of everyday diagnosis, from anemia to infection to clotting disorders, and a technician still does it by eye under a microscope: slow, tiring, subjective work that bounds a lab's throughput. This project teaches an off-the-shelf object detector to draw a box around every cell in a smear and label its type. The most consequential decision was what I chose not to do: no color, hue, brightness, or saturation augmentation during training. A blood cell's stain color is the label — the deep purple of a white cell's nucleus against the pale pink of red cells is exactly what a hematologist reads. Jitter the hue while holding the label fixed and you don't create a plausible new slide; you teach the model that the diagnostic signal is irrelevant. The full story of that reasoning lives in the writeup.

Torchvision RetinaNet, chosen deliberately as a single-stage detector: its focal-loss head down-weights the sea of easy background anchors so learning concentrates on hard cases, which is a genuine match for a microscopy field carpeted with hundreds of red cells against a handful of platelets — the extreme foreground/background and class imbalance (RBC ≫ WBC ≫ Platelet) that focal loss exists to tame. I took the raw BCCD in Pascal VOC format rather than a pre-augmented copy, precisely so I owned the augmentation policy instead of inheriting someone else's assumptions. Preprocessing: resize to 640×640, clip every box back inside image bounds, assert each lower bound strictly below its upper bound (degenerate boxes crash the pipeline), and keep box-less images as legitimate background. The only randomized augmentations were flips and rotations — both true invariances here, since a smear has no canonical left/right or up. Training config: 40 epochs, batch size 4 (VRAM-bound, not a research choice), learning rate 0.001 with no schedule, on an NVIDIA RTX 3050 Ti with 4 GB of VRAM. The premise was narrow and honest: how far does careful fine-tuning of a stock detector go, on modest hardware, without architecture surgery?

mAP 0.876 @ IoU 0.5 and 55.25% @ 0.50:0.95 on test, with Table I also reporting 88% accuracy at IoU 0.5. During training the best validation mAP at 0.50:0.95 landed around 58%, so the test number tracks what I saw in the loop rather than a lucky split. Set honestly against purpose-built detectors on the same benchmark, the fine-tuned RetinaNet did not win — YOLOv7 reaches 0.896, YOLOv5x 0.923, and CST-YOLO 0.927 @ IoU 0.5, a gap I don't paper over. What it did was land within a few points of task-specific architectures from pretrained weights on a 4 GB laptop, which is exactly the claim the premise set up. The honest limits are real too: red cells — the majority class — are the weak spot, because at BCCD's densities they overlap and clump and non-max suppression merges touching instances; and the model's color prior is pinned to BCCD's one staining protocol, so it generalizes poorly to a different lab's stain. The right fix there is real images from more protocols, not synthetic hue jitter — a different invariance claim entirely.

This was my proving ground for treating augmentation as a modeling decision, not a preprocessing checkbox — the "for each transform, does the true label survive it?" question I now run before enabling anything in production vision work, including validating vision models for kiosk monitoring at my job. What I'd change with more room: a learning-rate decay after epoch 30 to smooth the fluctuating late-training loss, more resolution and batch size on a bigger GPU, and — most of all — ablating the no-color policy rather than asserting it, running the color-augmented model as a measured control instead of a well-argued prior. One caveat kept me honest about the writeup itself: the prose has draft residue — the abstract and methodology mention blurring and brightness augmentations that the Training Details section (the ground truth: "random flips and rotations") contradicts, and one spot conflates the 88% and 0.876 figures. Because of those leftovers I keep the PDF available on request rather than treating it as polished.