bash - guest@parth - 80x24

guest@parth:~$ _

(increase audio for best effects)
PARTH_

Hi, welcome to my portfolio_

I am Parth Sharma, an AI/ML Engineer and Full-Stack Developer specializing in designing clinically reliable deep learning models and building high-performance web applications. I focus on bridging the gap between intelligent algorithms and production-ready software.

GITHUB LINKEDIN EMAIL
parth.ts
12345678 
export const developer = {
    name: "Parth Sharma",
    role: "AI/ML & Full-Stack Developer",
    focus: ["deep_learning", "fullstack"],
    available: true
};
AI/ML
PyTorch + TF
Dev
React + Node
Ops
Docker + REST
✓ Clinically reliable deep learning models
✓ Feature fusion architectures for AI diagnostics
✓ Scalable, real-time deterministic web systems
TypeScript
VIEW_PROFILE
Python ReactJS PyTorch NodeJS OpenCV Tailwind TensorFlow MongoDB Python ReactJS PyTorch NodeJS OpenCV Tailwind TensorFlow MongoDB
// SECTION: ABOUT_ME

SYSTEM
OVERVIEW.

I am an AI/ML-focused IT student with a strong foundation in building and deploying clinically reliable deep learning systems alongside scalable full-stack web applications.

My passion lies at the intersection of complex data modeling and intuitive user experiences. Whether it's crafting a responsive frontend or fine-tuning a neural network for medical imaging, I strive for efficiency, accuracy, and impact.

SYS.PRJ
0
// PROJECTS
SYS.PUB
0
// IEEE_PUB
SYS.INT
0
// INTERNSHIPS
SYS.GPA
0.0
// CGPA
// SECTION: CORE_SYSTEMS

CORE PROJECTS.

SYS.ID // 001

Skin Disease Classification

Trained a 7-class skin lesion classifier on the HAM10000 dataset using transfer learning. Deployed for real-time inference via REST APIs.

ResNet50 PyTorch React Node.js
VIEW_PROJECT →
SYS.ID // 002

Diabetic Foot Ulcer Classifier

CNN-based image classification system for automated detection and grading of diabetic foot ulcers with preprocessing for model generalization.

Python CNN OpenCV NumPy
VIEW_PROJECT →
SYS.ID // 003

Doctor Appointment System

Full-stack platform with JWT authentication, dynamic dashboards for roles, and robust appointment lifecycle logic.

MERN Stack JWT MongoDB
VIEW_PROJECT →
SYS.ID // 004

Real-Time Chat App

Bidirectional messaging system supporting private and group chats, scalable WebSocket events, and persistent storage.

MERN Stack Socket.io React.js
VIEW_PROJECT →
// SECTION: RESEARCH_PUB
> IEEE_PUB
// RESEARCH PAPER

A Deep Learning Framework for Automated Multi-Class Skin Disease Detection Using Dermoscopic Images

Joshin K Saju, Deepayan Thakur, Parth Sharma

ICAISS 2025 | Amity University

Abstract

Skin diseases form a major world health concern, where early detection is crucial for improving patient outcomes and survival rates. However, skin lesion similarity across categories and severe class imbalance create a complex multi-class classification problem. A new effective framework based upon deep learning is presented in this research paper, targeting automated dermoscopic image analysis.

Transfer learning-based CNNs, pretrained on ImageNet, are utilized to analyze skin lesion characteristics across seven diagnostic categories. Preprocessing methods, including resizing to uniform resolution, intensity normalization to [0,1], and data augmentation through rotation, flipping, and color jittering, are utilized to achieve better reproducibility and model robustness.

The training minimizes false negatives for malignant classes by using categorical cross-entropy loss along with the Adam optimizer, incorporating sensitivity-driven learning with class-weighted sampling. The proposed system achieves 89% overall accuracy and 99% melanoma recall on 2003 held-out test images, demonstrating strong clinical viability.

Keywords

Skin disease detection Deep learning CNN Transfer learning HAM10000 Dermoscopy
1

I. Introduction

Human skin functions as a primary defense system against environmental hazards. Despite its accessibility for visual examination, skin disease diagnosis remains difficult due to similar physical presentations across disorders and the subjective nature of clinical assessment. Malignant skin cancers like melanoma carry high mortality rates when detection is delayed, with survival dropping sharply beyond Stage II.

Traditional diagnostic methods rely heavily on dermatologist expertise and face poor generalization due to variations in illumination, skin tones, imaging devices, and lesion complexities. The emergence of CNNs has marked a breakthrough in dermatology image analysis, enabling automated feature extraction that surpasses handcrafted descriptors. The HAM10000 dataset, containing over 10,000 dermoscopic images across seven categories, provides a standardized benchmark for evaluating such systems.

Challenges

  • Limited availability of large, diverse, and well-annotated datasets across all skin tones and demographics.
  • Severe class imbalance in multi-class classification — melanocytic nevi dominating with ~67% of samples.
  • High visual similarity among different diseases (e.g., benign keratosis vs. actinic keratosis, eczema vs. psoriasis).
  • Image quality variations caused by inconsistent lighting, noise, hair occlusion, and color calibration differences.
  • Clinical requirement for high sensitivity in malignant classes, where false negatives carry life-threatening consequences.
2

III. Methodology

The system evaluates seven diagnostic categories from the HAM10000 dataset: benign keratosis-like lesions (bkl), melanocytic nevi (nv), dermatofibroma (df), melanoma (mel), vascular lesions (vasc), basal cell carcinoma (bcc), and actinic keratosis (akiec). The dataset was split into 80% training and 20% testing subsets with stratified sampling to preserve class distribution.

Data Prep: Images resized to 224×224 uniform resolution. Intensities normalized to [0,1] using min-max normalization to stabilize gradient updates. Augmentation strategies including random rotation (±30°), horizontal/vertical flips, and color jittering were applied to minority classes to mitigate imbalance.
Feature Extraction: Employs deep hierarchical extraction using pretrained CNNs (ResNet50 backbone). Transfer learning on ImageNet weights enables exploitation of low-level edges, textures, and high-level semantic representations without training from scratch.
Optimization: Categorical cross-entropy loss with Adam optimizer (lr=0.0001). Dropout (p=0.5) randomly deactivates neurons to reduce overfitting. Learning rate scheduling with ReduceLROnPlateau monitors validation loss for adaptive convergence.
Architecture: Final classification head consists of Global Average Pooling → Dense(512, ReLU) → Dropout → Dense(7, Softmax). Batch normalization layers stabilize intermediate activations during training.
3

IV. Results

89%_
Overall Accuracy
99%_
Melanoma Recall

System demonstrated strong generalization on 2003 held-out test images. Multi-class ROC-AUC values ranged from 0.98 to 1.00, confirming excellent discriminative ability across all seven categories. Per-class F1 scores exceeded 0.85 for majority classes.

The confusion matrix exhibits strong diagonal dominance. Error analysis revealed misclassifications occurred primarily between benign keratosis and actinic keratosis — visually similar lesion types — generating clinically safer errors that prioritize sensitivity over specificity.

Notably, the model maintained >95% sensitivity for both melanoma and basal cell carcinoma, the two most clinically dangerous categories. Precision-recall curves further validated that the framework avoids dangerous false-negative predictions in high-risk diagnostic scenarios.

4

V. Conclusion

The framework achieves a strong balance between accuracy, sensitivity, and robustness across all seven diagnostic categories. The exceptionally high recall for malignant lesions — 99% for melanoma and >95% for basal cell carcinoma — highlights its suitability as a clinical decision-support tool.

Future work includes expanding the dataset to encompass diverse skin tones, integrating attention mechanisms for improved interpretability, and deploying the model as a real-time mobile diagnostic aid for resource-limited clinical settings.

READ FULL PAPER
PAGE 1 / 6
// LOG: EXPERIENCE_VAULT

EXPERIENCE VAULT.

// Click a notebook

SYS.EXP // 001

AI/ML & Web Dev Intern

> IHFC, IIT Delhi

INDEX_REF // 001

// 2 Months · Offline
SYS.EXP // 002

Web Dev Intern

> GIIT Solutions

INDEX_REF // 002

// May–Jun 2025
SYS.EXP // 003

Web Dev Intern

> Afame Technologies

INDEX_REF // 003

// May–Jun 2024 · Remote
SYS.EXP // 004

Freelance Photo & Video Editor

> Independent

INDEX_REF // 004

// Ongoing
// LOG: ACADEMIC_RECORD

EDU CATION.

README
// SUMMARY — README.EXE
v1.0.0
// PROCESS: IDENTITY

Parth Sharma

B.Tech Information Technology student · AI/ML researcher · Full-Stack developer · Photo & Video editor. Building at the intersection of intelligence and interface.

STATUS: OPEN TO WORK
2022 – 2026 · Delhi
01
ABOUT
ABOUT
> Clinically reliable AI systems · open-source contributor · hackathon finalist · multi-disciplinary creator.
02
PROJECTS
PROJECTS
> Full-stack web apps · deep learning pipelines · IoT systems · browser extensions. Built to ship.
03
RESEARCH
RESEARCH
> Medical AI · dermatology classification · retinal OCT · published IEEE paper (IEEE Xplore, 2024).
04
EXPERIENCE
EXPERIENCE
> IIT Delhi (research intern) · Curo Wealth (dev intern) · Freelance photo & video editor.
05
EDUCATION
EDUCATION
> B.Tech IT · Sharda University · CGPA 7.5. XII CBSE 65% · X CBSE 70% · Don Bosco School.
06
STACK
STACK
> Python · TypeScript · React · Next.js · PyTorch · TensorFlow · Firebase · FastAPI · Tailwind.
07
EMAIL
MAILTO
parthsharmaro2@gmail.com
08
LINKEDIN
CONNECT
linkedin.com/in/pogth
MODULES LOADED: 8 STACK: FULL AVAILABILITY: OPEN LOCATION: DELHI AGE: —