Ultrasound enters
A transabdominal or transvaginal image is standardized to a 512 × 512 model input.
AI-ASSISTED OVARIAN ULTRASOUND
Detect suspicious ovarian findings with visual precision.
A presentation-ready object-detection workflow that identifies and localizes eight ovarian ultrasound categories using an EfficientDet-D0 backbone and augmented 512 × 512 imaging data.
Powered with ❤️ by Solicitous ®
8Ultrasound categories
512²Model input resolution
14KInspected annotations
D0EfficientDet backbone
REAL-TIME ULTRASOUND REVIEW
MODEL READY
Active categoryTeratoma candidate
92%confidence
OutputBounding box + class
THE CORE IDEA
What is actually happening?
The system is not drawing a full pixel mask. It performs object detection: it finds a suspicious region, places a box around it, assigns one of eight categories, and reports a confidence score.
Features are fused
EfficientDet combines multi-scale visual features so both large cystic areas and smaller focal findings can be considered.
Candidate boxes emerge
The detector proposes regions and suppresses overlapping duplicates through non-maximum suppression.
Class + confidence
Each retained region receives a category label and confidence; the demo threshold is configurable around 0.40.
Important: this is an experimental research workflow, not a diagnostic device. The repository itself reports uneven class performance, so clinical validation and expert review are essential.
ANIMATED DATASET THEATRE
Eight-class detection, one controlled viewer.
Play through representative model outputs. Every ultrasound is displayed with contain-fit framing, so the complete source image remains visible.
Detected class
Chocolate cyst
A blood-filled endometriotic cyst candidate localized by the detector.
58%
Model confidence on this exampleEIGHT-CATEGORY TAXONOMY
What the category numbers mean.
The source outputs use numeric IDs. This presentation translates them into readable clinical category names.
Chocolate cyst
Endometrioma-like cystic finding.
3,136 training objectsSerous cystadenoma
Typically benign epithelial cystic tumor.
2,212 training objectsTeratoma
Mixed echogenic appearance candidate.
3,080 training objectsTheca cell tumor
Sex-cord stromal tumor candidate.
756 training objectsSimple cyst
Thin-walled, fluid-filled cystic pattern.
812 training objectsNormal ovary
Non-tumor ovarian appearance.
2,492 training objectsMucinous cystadenoma
Multiloculated mucinous-pattern candidate.
868 training objectsHigh-grade serous
High-risk malignant-pattern candidate.
644 training objectsBEFORE / AFTER DETECTION
See exactly what the model adds.
Drag the divider. The left side is the original ultrasound; the right side is the model output with its localized region and label.
↔
ORIGINALAI OUTPUT
MODEL PIPELINE
From augmented scan to saved detector.
01OTU_2D ultrasound + masksEight-category source data
→
02Rotate, flip, standardize512 × 512 augmented master set
→
03YOLO annotationsBoxes derived from masks
→
04TFRecord + COCOTrain/validation and test formats
→
05EfficientDet-D0Fine-tuned from a COCO checkpoint
→
06SavedModel inferenceBoxes, class IDs, confidence
DATASET INSPECTOR
A deliberately animated source-image wall
The moving viewport scans across the repository’s annotated training contact sheet. It demonstrates the variety of acquisition angles, lesion sizes, probe shapes, grayscale patterns and bounding-box scales presented to the model.
- 14,000 inspected training objects across eight classes
- Strong imbalance between common and rare categories
- Augmentation increases orientation and appearance diversity
RESEARCH EVIDENCE
Performance shown honestly, not cosmetically.
The reported numbers indicate a useful experimental prototype, but also clear headroom before clinical deployment.
Validation epoch 540.401mAP [0.50:0.95]
Validation epoch 540.560mAP @ 0.50 IoU
Held-out test0.294mAP [0.50:0.95]
Held-out test0.395mean average recall
Validation F-score and COCO mAP across 54 epochs.
WHY THE RESULTS VARY
Class imbalance is visible in the benchmark.
At epoch 54, AP ranges from about 0.60 for class 0 to about 0.06 for class 4. That gap can reflect limited examples, overlapping ultrasound appearances, annotation variation and model capacity. The right next step is not to hide the weakness—it is to expand balanced data, add expert review and validate on external scanners.
PRESENTATION VALUE
How this can support an ultrasound-instrument story.
On-console second look
Surface candidate regions after image capture while keeping the clinician in control.
Standardized documentation
Store box coordinates, predicted category and confidence with each review case.
Teaching and QA mode
Compare model output with expert annotations for training and retrospective quality review.
Data flywheel
Capture expert corrections to build a better, balanced, scanner-diverse validation dataset.
OVARIAN ULTRASOUND · OBJECT DETECTION · RESEARCH DEMO
Return to top ↑
Clear localization. Transparent evidence. Better conversations.
OvaScope AI turns a complex detection repository into a polished, understandable presentation experience.