Page 65 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Technology TinyML, Smartphone Microscopy, Malaria Detection, Deep Learning, Edge AI,
Keywords Trochozoite Identification, WBC Detection 4.1-Healthcare
Data Availabil- Public [1]
ity

Metadata (Type Visual data (high-resolution images of blood slides)
of Data)

Model Training The AI model is trained using supervised learning techniques, specifically deep
and Fine-Tun- learning with convolutional neural networks (CNNs), to classify and detect
ing malaria parasites (Trochozoites) and WBCs from blood slide images. Fine-tun-
ing is performed using labeled visual data from the Lacuna Malaria Detection
Challenge dataset to improve accuracy and adapt the model for real-world
variability in image quality. TinyML frameworks are then used to optimize the
model for efficient deployment on resource-constrained smartphone devices.

Testbeds or Lacuna Malaria Detection Challenge [2]
Pilot Deploy-
ments
Code reposito- Smartphone-over-Microscope-Diagnosis [3]
ries

2 Use Case Description

2�1 Description

This use case centers on utilizing a smartphone mounted over a microscope to improve the
diagnosis of malaria parasites in rural communities. The approach taps into TinyML's capabilities
for real-time image analysis, thereby empowering community health workers with a portable
and efficient diagnostic tool. By leveraging a smartphone to capture high-resolution images
of blood slides through the microscope eyepiece, the system aims to facilitate accurate and
timely detection of malaria infections, specifically identifying Trochozoites and WBCs.

The problem this use case addresses is limited access to reliable and timely malaria diagnosis in
remote areas, where resources and trained professionals may be scarce. Current solutions, such
as traditional microscopy or RDTs, often face limitations. Conventional microscopy requires
skilled technicians and may not be feasible in rural settings, while RDTs can yield false negatives
and are less accurate.

The dataset employed for training the model is sourced from the Lacuna Malaria Detection
Challenge, which contains diverse and labeled images of blood slides, essential for calibrating
the algorithm for effective detection. Utilizing this dataset ensures a robust model capable of
distinguishing between the relevant classes—Trochozoites and WBCs.

The benefits of this proposed solution include enhanced accessibility to malaria diagnosis,
reduced dependency on laboratory facilities, and the utilization of familiar technology
(smartphones) by healthcare workers. However, potential drawbacks include the need for
adequate smartphone hardware and battery life, which could affect performance in areas
with unreliable electricity supply. Additionally, while the model can achieve high accuracy with

29