Liverscan++: raspberry pi-based web-integrated device for detection and correlation of focal liver lesions from triphasic contrast-enhanced ct scans using deep learning algorithm/ Jowan Clyderie C. Anciano, Chelsea R. Angeles, Gillian C. Gunita, Ann Jenette M. Gutierrez, Desiree Ann B. Nieto, and Mikko O. Reyes.--

Bachelor's thesis

College of Engineering.-- Bachelor of science of electronics engineering: Technological University of the Philippines, 2025.

Includes bibliographic references and index.

The detection of results involving liver diseases requires manual interpretation that
often takes several days to weeks to be completed leading to delays that may result to late
treatment options for the patients. The LIVERSCAN++ system, a Raspberry Pi-based,
web-integrated device, was developed to detect and correlate focal liver lesions from
triphasic contrast-enhanced computer tomography (CECT) images using a deep learning
algorithm. The automated detection involves a pre-processing technique with the use of
three phases of gray scaled CECT scans, namely unenhanced, arterial, and portal venous,
by converting it into RGB format through assigning each phase to a specific channel and
stacking it vertically. The stacked images underwent custom training using YOLOv11
Instance Segmentation model to enable the system to identify specific liver lesions
accurately and correlate each phase of the CECT scans’ phases. Furthermore, the device is
paired with a graphical user interface that was designed solely for medical professionals’
use. This interface was created for uploading images, accessing patients’ database, and
making predictions based on the automated results. The device only acts as an aid for
professionals to ease clinical workloads and improve radiological assessment. A website
platform was integrated with the interface and was created for both medical professionals
and their patients. Patients can access their verified results in real time. The model achieved
strong performance with a mean Average Precision (mAP@50) of 0.902 and overall
accuracy of 96.7%. Processing times averaged 6.5 seconds over Wi-Fi and 10.1 seconds
on mobile data. Usability testing showed high effectiveness (score 4.90 on ISO 62366),
while software quality evaluation (ISO 25010) indicated room for improvement, especially
by expanding the training data and refining validation to boost accuracy and reliability.