Technology

Models & architecture

Explore the models behind each phase of the pipeline, from the 7 model ensemble to the 11 class object detector.

Phase 1 models

Phase 2 classifier

Phase 3 classes

Phase 1: Fecal detection ensemble

Seven TensorFlow / Keras architectures, each fine tuned independently on the same fecal detection dataset. Their outputs are combined through majority voting to produce a single fecal vs non fecal decision.

VGG19

Deep CNN

19 weight layers, well studied baseline for transfer learning with strong feature extraction.

ResNet50

Residual Network

50 layer residual network with skip connections that prevent vanishing gradients.

DenseNet169

Dense Blocks

169 layer network where every layer is connected to every other, maximizing feature reuse.

EfficientNetB0

Compound Scaling

Balanced depth, width, and resolution scaling for best accuracy per FLOP.

MobileNetV2

Lightweight CNN

Inverted residuals and linear bottlenecks, optimized for mobile and edge deployment.

NASNetMobile

NAS optimized

Architecture discovered via neural architecture search, tuned for mobile scale compute.

ConvNeXtBase

Modern CNN

A pure ConvNet that matches vision transformers by modernizing classic design choices.

View on Hugging Face

Phase 2: Helminth screening

Binary classifier: Helminth vs Non Helminth

A single dedicated model receives confirmed fecal slides and determines whether parasitic helminth eggs or organisms are present. This binary gate prevents the heavier object detection model from running on clean samples, improving both speed and specificity.

Phase 3: Species identification

11 class object detection with bounding boxes

Helminth positive slides are scanned by an object detection model trained to localize and classify eggs or organisms from 11 parasitic species. Each detection includes a bounding box drawn directly on the microscopy image, along with a species label and confidence score.

Detectable species

Ascaris lumbricoidesCapillaria philippinensisEnterobius vermicularisFasciolopsis buskiHookworm eggHymenolepis diminutaHymenolepis nanaOpisthorchis viverrineParagonimus sppTaenia spp. eggTrichuris trichiura

Interactive diagram

How ensemble voting works

Each Phase 1 model independently votes. A simple majority (≥4 of 7) determines the outcome.

VGG19Fecal

ResNet50Fecal

DenseNet169Fecal

EfficientNetB0Fecal

MobileNetV2Fecal

NASNetMobileNon fecal

ConvNeXtBaseNon fecal

Majority vote: Fecal

5 of 7 models agree · Advance to Phase 2

See the pipeline in action

Create a free account, upload a microscopy slide, and watch each phase run in real time.

Get started free Learn more