CervixAI Botswana — Automated Cervical Cancer Detection

VIA Image Input

Sample — Normal cervix

Select Sample Image

Classification Result

Preprocessing image...

Select a sample image and click Analyze to run the ResNet-50 classifier

Model Architecture

ArchitectureResNet-50

Transfer LearningImageNet pretrained

Input Size224 × 224 px

Classes3 (Normal · Pre · Cancer)

AugmentationRotation, flip, contrast

Validation5-fold cross-val

Dataset Split

70%

Training

2,100 imgs

15%

Validation

450 imgs

15%

Test

450 imgs

91.4%

Sensitivity

↑ 6.4% above target

88.7%

Specificity

↑ 3.7% above target

0.923

AUC-ROC

Strong discrimination

89.9%

F1-Score (macro)

Precancerous: 87.2%

Confusion Matrix — Test Set (n=450)

	Pred: Normal	Pred: Precancerous	Pred: Cancer
True: Normal	182	14	4
True: Precancerous	12	141	9
True: Cancer	2	7	79

Per-class F1 Scores

Normal

F1 = 0.921

92.1

Precancerous

F1 = 0.872

87.2

Cancer

F1 = 0.904

90.4

5-Fold Cross-Validation AUC

Fold 1

0.918

Fold 2

0.926

Fold 3

0.913

Fold 4

0.931

Fold 5

0.928

Mean AUC: 0.923 ± 0.007

Referral Time Reduction

AI-assisted

42 days avg

Standard care

76 days avg

Reduction: 34 days (44.7% faster · H2 confirmed)

Patient Queue — Tlokweng PHC

Select a patient to view referral timeline

Study Context

Cervical cancer is a leading cause of cancer mortality among women in Botswana, with most cases diagnosed at advanced stages. This prototype demonstrates a ResNet-50 CNN system for VIA image classification in rural PHC clinics, designed to support clinician decision-making without replacing specialist care.

Research Questions

RQ1 — Can ResNet-50 accurately classify VIA images into normal, precancerous, or suspected cancer categories?

RQ2 — Does CNN-assisted tele-screening reduce average referral completion time?

RQ3 — How do patients and PHC staff perceive usability of AI-assisted screening?

RQ4 — Is CNN-assisted tele-screening feasible in low-resource rural PHC clinics?

Study Hypotheses

H1 — ResNet-50 achieves sensitivity ≥85% and specificity ≥85% vs expert assessment

H2 — AI tele-screening reduces referral time by ≥30 days vs standard care

H3 — Patients and staff report >80% positive feedback on usability

ML Pipeline

VIA Capture

→

Resize 224×224

→

Normalize

→

Augment

→

ResNet-50

→

3-class output

Study Parameters

Participants500 women

Total Images3,000 VIA

PHC Clinics10 rural sites

Age Range25–49 years

Training split70 / 15 / 15 %

Validation5-fold + external

Expected Impact

Anticipated to increase early detection sensitivity by at least 15% and reduce referral time by ≥30 days. Scalable to rural PHC clinics across Botswana and similar LMIC settings.

Early detection↑ 15% sensitivity