AI CASE STUDY / MEDICAL INTELLIGENCE

DERMTRACK

Intelligent Skin Health Monitoring

A Hybrid CNN-RNN Framework for Longitudinal Tracking and Progression Analysis of Dermatological Conditions
DEVELOPER DHEERAJ KOLLI
ROLE AI / ML • Computer Vision • Full-Stack
DATASET HAM10000
MODEL MobileNetV2 + LSTM
STAGE // HYBRID AI LONGITUDINAL PIPELINE ● LIVE NEURAL SIMULATION
🖼️ IMAGE CAPTURE
➔
🧠 MOBILENETV2 (CNN)
➔
📊 1280-D FEATURE VECTOR
➔
⏱️ LSTM (RNN) TEMPORAL
➔
📈 PROGRESSION INSIGHT
01 — PROJECT OVERVIEW

BEYOND ISOLATED SNAPSHOTS

DermTrack is an AI-powered application designed for longitudinal tracking of dermatological conditions. The central idea is to move beyond analyzing a single skin image as an isolated event.

01

IMAGE DATA

High-resolution spatial feature extraction captured over periodic intervals.

02

TIME / SEQUENCES

Sequential temporal modeling tracking condition dynamics across time.

03

SYMPTOM LOGGING

Patient-reported metrics such as pain and itch intensity integrated into temporal analysis.

02 — THE PROBLEM

CHALLENGES IN SKIN CONDITION MONITORING

01

SUBJECTIVE TRACKING

Patients struggle to accurately describe changes in their skin conditions over time, relying heavily on memory and perception.

02

SPORADIC DATA

Clinical visits provide isolated points of observation and may miss critical progression changes between appointments.

03

LACK OF OBJECTIVITY

Without longitudinal information, determining whether a condition is improving, worsening, or remaining stable is difficult.

TRADITIONAL APPROACH FROM ISOLATED SNAPSHOTS
↓
DERMTRACK APPROACH TO A LONGITUDINAL PROGRESSION STORY
03 — THE SOLUTION

THE DERMTRACK APPROACH

A mobile AI-powered platform for data-driven skin health progression tracking.

01

LOG

Users capture periodic photos of their condition and log associated symptoms such as itch and pain.

02

ANALYZE

The hybrid AI model analyzes visual information and temporal sequences.

03

VISUALIZE

The system generates a data-driven timeline representing progression states: IMPROVING, WORSENING, or STATIC.

04 — CORE CONCEPT

IMAGE + TIME + SYMPTOMS → PROGRESSION INSIGHT

CNN

SPATIAL INTELLIGENCE

"What is visible in the image?"

+
LSTM / RNN

TEMPORAL INTELLIGENCE

"How is it changing over time?"

=
DERMTRACK

LONGITUDINAL INSIGHT

Combines spatial and temporal intelligence.

05 — SYSTEM ARCHITECTURE

END-TO-END TECHNICAL ARCHITECTURE

IMAGE CAPTURE Periodic Photos
IMAGE PREPROCESSING Standardization & Resizing
MOBILENETV2 (CNN) Spatial Transfer Learning
1280-D FEATURE VECTOR Visual Embeddings
+ SYMPTOM METADATA (Pain / Itch Logs)
LSTM / RNN Sequence Modeling
TEMPORAL ANALYSIS Longitudinal Trends
PROGRESSION INSIGHT Improving / Static / Worsening
06 — NEURAL MODULES

SPATIAL & TEMPORAL INTELLIGENCE

01

SPATIAL INTELLIGENCE

CNN / MobileNetV2

A CNN processes individual dermatological images and extracts visual features. MobileNetV2 is utilized for transfer learning, outputting 1,280-dimensional feature vectors per image frame.

IMAGE ➔ MOBILENETV2 ➔ 1280-D VECTOR
02

TEMPORAL INTELLIGENCE

LSTM / RNN

An RNN using LSTM cells models temporal progression over sequences of image-derived feature vectors and associated symptom metadata across multiple time steps.

IMAGE SEQUENCE ➔ LSTM CELL ➔ TEMPORAL PATTERN
07 — DATASET & PIPELINE

DATASET & PREPROCESSING

DATASET IDENTIFIER: HAM10000

Multi-source dermatoscopic images utilized for training spatial feature representations and verifying feature extraction pipelines.

PREPROCESSING PIPELINE
RAW IMAGE
➔
STANDARDIZATION & RESIZING
➔
PIXEL NORMALIZATION
➔
DATA AUGMENTATION
➔
MODEL INPUT
08 — TECH STACK

TECHNOLOGIES & TOOLS

PYTHON PYTORCH MOBILENETV2 CNN RNN LSTM OPENCV FLASK
09 — EVALUATION

MODEL EVALUATION & METRICS

89% CNN CLASSIFICATION ACCURACY
92% LSTM TEMPORAL PRECISION
87% ENSEMBLE MODEL PERFORMANCE
* Reported project evaluation metrics; not a substitute for clinical validation.
10 — COMPARISON

FROM STATIC CLASSIFICATION TO LONGITUDINAL ANALYSIS

STATIC APPROACH

  • Single image snapshot
  • Isolated classification
  • One-time result without history
DERMTRACK APPROACH

LONGITUDINAL AI

  • Image sequence over time
  • Patient-reported symptom integration
  • CNN spatial + LSTM temporal modeling
  • Dynamic progression trend insight
11 — CREDITS

MY ROLE

DHEERAJ KOLLI

AI / ML • Computer Vision • Full-Stack Development

"Designed and developed DermTrack as a hybrid AI-powered approach for longitudinal skin-condition monitoring."

12 — ROADMAP

FUTURE DEVELOPMENT

01

EXPANDED CONDITIONS

Support a wider range of dermatological conditions.

02

PREDICTIVE ANALYTICS

Explore predictive models for flare-ups or treatment non-response.

03

EHR INTEGRATION

Potential integration with Electronic Health Records.

04

MOBILE APP EXPANSION

Further development of the native mobile application experience.

05

WEARABLE SENSOR DATA

Potential integration of wearable sensor telemetry.