codeshujaaa/kenyanmalarai-detect

Plasmodium Life Stage Detection on Thin Blood Smear using YOLO26m

This model detects and classifies Plasmodium falciparum life stages in Giemsa-stained Thin blood smear images using a YOLO26m object detection architecture. The three target classes are Ring, Trophozoite, and Schizont the three intraerythrocytic stages observable under light microscopy.

The work is motivated by the clinical need for automated malaria staging in resource-limited settings, where trained microscopists are not always available and diagnostic turnaround time is critical.

This repository documents the full development journey from V1 to V2, including the challenges encountered, decisions made, and improvements achieved at each stage.

V1 Dataset

The dataset consists of Giemsa-stained blood smear images annotated with bounding boxes and life stage labels at the individual cell level. Annotations were sourced from an Excel file mapping image filenames to cell center coordinates and stage labels.

Original Label	Mapped Class	Split	Ratio	Annotations	Class	Crop Size
R, LR-ET	Ring	Train	70%	Ring oversampling x3 applied	Ring	80 px
LT, MT	Trophozoite	Val	15%	122 images, 439 instances	Trophozoite	110 px
Esch, Lsch, Seg	Schizont	Test	15%	122 images, 436 instances	Schizont	150 px

---

V1 Model Details

Property	Value	Property	Value
Architecture	YOLO26m	Parameters	20,351,765
Framework	Ultralytics 8.4.23	GFLOPs	67.9
Input Size	512 x 512 px	Training Device	Tesla P100 16GB
Classes	Ring, Schizont, Trophozoite	Epochs	80
Optimizer	AdamW	Learning Rate	5e-5 (cosine decay to 0.001)

source code can be found : https://github.com/codeshujaa/Plasmodium-Life-Stage-Detection/blob/main/README.md

V1 Training

Training was conducted with strong data augmentation to compensate for the limited dataset size.

The classification loss weight was increased to 3.0 to address the morphological ambiguity between transitional stages, particularly between Trophozoite and the adjacent Ring and Schizont classes.

V1 Training Curves

The TensorBoard training curves are included in this repository under runs/plasmodium/. To visualise them locally

tensorboard --logdir runs/plasmodium

Key observations from the curves

Loss/train_box converged from ~1.2 to 0.81 by epoch 79

Loss/train_cls converged from ~12 to 6.95 — the highest absolute loss, reflecting the classification difficulty of morphologically similar stages

Loss/val_box and Loss/val_cls both decreased consistently without divergence, confirming the model did not overfit

mAP/mAP50 reached 0.693 at epoch 79 with a smooth convergence curve

mAP/mAP50_95 reached 0.494 at epoch 79

Both mAP curves plateau after epoch 60, indicating the model reached the limit of what the current dataset volume supports.

---

V1 Evaluation Results

Inference was run at conf=0.15 and iou=0.5 on the held-out test set.

Detection Performance on Test Set V1

Class	Precision	Recall	mAP@0.5	mAP@0.5:0.95
All	0.666	0.753	0.735	0.513
Ring	0.623	0.586	0.551	0.377
Schizont	0.733	0.937	0.917	0.646
Trophozoite	0.641	0.737	0.737	0.516

Classification Performance on Test Set V1

Class	Precision	Recall	F1	Support
Ring	0.68	0.57	0.62	115
Schizont	0.70	0.76	0.73	214
Trophozoite	0.42	0.42	0.42	105
Accuracy			0.63	434
Macro avg	0.60	0.58	0.59	434
Weighted avg	0.63	0.63	0.62	434

---

Confusion Matrix V1

The confusion matrix

Key observations

• Schizont is the most reliably detected class, achieving 0.917 mAP@0.5, consistent with its large and morphologically distinct appearance
• Ring stage shows the highest background confusion — small cell size and light staining intensity make Ring cells the hardest to detect at this dataset scale
• Trophozoite is confused in both directions, with some predictions falling toward Ring and others toward Schizont, reflecting its intermediate morphological position in the parasite life cycle

---

Sample Detections

Sample detection outputs on validation images.

---

V1 How to Use

!pip install ultralytics
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Load V2 model
weights_path = hf_hub_download(
    repo_id  = "codeshujaaa/kenyanmalarai-detect",
    filename = "weights/best.pt"
)

model   = YOLO(weights_path)
results = model.predict(
    source = "https://en.pimg.jp/023/550/618/1/23550618.jpg",
    conf   = 0.15,
    iou    = 0.5,
    imgsz  = 640,
)

results[0].show()

V2 Data Expansion

Based on V1 findings, a second annotated dataset was sourced from Roboflow Universe and merged with the original dataset. The Roboflow dataset contained seven classes. Only Ring, Schizont, and Trophozoite annotations were retained and remapped to match the V1 class ID schema before merging.

Data growth from V1 to V2

Class	V1 Instances	V2 Instances	Growth
Ring	127	1356	10x
Schizont	214	1546	7x
Trophozoite	95	2283	24x

Training was extended from 80 to 120 epochs and image size was increased from 512 to 640. Mosaic augmentation was disabled for the final 10 epochs to allow the model to consolidate detection on full-size images.

V2 Model Details

Property	Value	Property	Value
Architecture	YOLO26m	Parameters	20,351,765
Framework	Ultralytics 8.4.24	GFLOPs	67.9
Input Size	640 x 640 px	Training Device	Tesla T4 x2
Classes	Ring, Schizont, Trophozoite	Epochs	120
Optimizer	AdamW	Learning Rate	5e-5 (cosine decay)

V2 Training Curves

V2 Detection Performance on Test Set

Class	Precision	Recall	mAP@0.5	mAP@0.5:0.95
All	0.747	0.812	0.830	0.638
Ring	0.698	0.701	0.735	0.565
Schizont	0.775	0.930	0.933	0.750
Trophozoite	0.770	0.804	0.822	0.601

V2 Classification Performance — Test Set

Class	Precision	Recall	F1	Support
Ring	0.71	0.72	0.72	175
Schizont	0.69	0.79	0.74	232
Trophozoite	0.80	0.70	0.74	317
Accuracy			0.73	724
Macro avg	0.73	0.74	0.73	724
Weighted avg	0.74	0.73	0.74	724

V2 Confusion Matrix

V2 Sample Detections

---

V1 to V2 — Progress

Metric	V1	V2	Improvement
Overall mAP@0.5	0.735	0.830	+0.095
Overall mAP@0.5:0.95	0.513	0.638	+0.125
Ring mAP@0.5	0.551	0.735	+0.184
Trophozoite F1	0.42	0.74	+0.32
Overall accuracy	0.63	0.73	+0.10

The most significant gain was Trophozoite F1 improving from 0.42 to 0.74 a direct result of increasing its training instances from 95 to 2283. This confirmed the V1 hypothesis that data volume was the primary bottleneck.

---

Limitations

The primary limitation of V1 model is dataset size. Instance counts across classes are uneven, with Trophozoite having the fewest examples (95 instances in the test split). The direct relationship between instance count and F1 score observed across all three classes confirms that further performance gains require additional annotated data rather than architectural changes.

In V2 Ring stage remains the weakest class at mAP@0.5 of 0.735. Ring cells are the smallest and most morphologically subtle stage. Further gains require additional annotated Ring stage images.

The model is a research prototype and is not validated for clinical deployment.

---

---

V2 How to Use

!pip install ultralytics
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Load V2 model
weights_path = hf_hub_download(
    repo_id  = "codeshujaaa/kenyanmalarai-detect",
    filename = "weights/best_v2.pt"
)

model   = YOLO(weights_path)
results = model.predict(
    source = "https://media.istockphoto.com/id/2169552309/photo/plasmodium-vivax-in-thin-film-under-microscope.jpg?s=612x612&w=0&k=20&c=ZjxGlaMKoPQEvJyRAwlZWgugPZMNIV1EiJeO1oP5Z8A=",
    conf   = 0.15,
    iou    = 0.5,
    imgsz  = 640,
)

results[0].show()

---

Repository Structure

kenyanmalarai-detect/
    weights/
        # V1 model
        best.pt
        # V2 model          
        best_v2.pt      
    runs/
        # V1 TensorBoard logs
        plasmodium/
        # V2 TensorBoard logs     
        plasmodium_v2/   
    README.md

Citation

If you use this model or findings from this work in your research, please cite this repository.

@misc{mwangi2025plasmodium,
  author       = {Denis Mwangi},
  title        = {Plasmodium Life Stage Detection Using YOLO26m on Giemsa-Stained Blood Smears},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/kenyanmalarai-detect}}
}    

Acknowledgements

Training infrastructure provided by Kaggle