update: Add detalhes do paper

Dockerfile

@@ -9,7 +9,9 @@ ENV PYTHONDONTWRITEBYTECODE=1 \
     PATH=/home/user/.local/bin:$PATH \
     HF_HOME=/tmp/.huggingface \
     TRANSFORMERS_CACHE=/tmp/.cache/huggingface/transformers \
-    HUGGINGFACE_HUB_CACHE=/tmp/.cache/huggingface/hub
+    HUGGINGFACE_HUB_CACHE=/tmp/.cache/huggingface/hub\
+    PAPER_URL=https://ieeexplore.ieee.org/Xplore/home.jsp\
+    PAPER_TITLE="RG-DermNet: A Multimodal Attention-Based Model with Residual Block Usage for Skin Lesion Classification \n   PAPER_DESCRIPTION=Skin cancer accounts for nearly one-third of all diag-\nnosed tumors worldwide, making early and accurate recognition\ncritical for improving patient outcomes. In this work, we pro-\npose RG-DermNet, a multimodal deep learning framework that\nintegrates skin lesion images with structured clinical metadata\nthrough a residual gated-attention (RG-ATT) fusion mechanism.\nThe architecture combines CNN- and Transformer-based visual\nbackbones with a lightweight one-hot encoding pipeline for meta-\ndata, enabling effective cross-modal interaction. The proposed\nmodel is evaluated using a patient-wise cross-validation protocol\nacross four dermatological datasets with heterogeneous metadata.\nOn PAD-UFES-20, using Caformer-B36 as the visual backbone,\nRG-DermNet achieves an accuracy of 0.75 ± 0.05, balanced\naccuracy of 0.78 ± 0.03, F1-score of 0.77 ± 0.04, and AUC of\n0.95 ± 0.01, outperforming existing multimodal baselines under\nthe same evaluation setting. In addition, a SHAP-based analysis\nprovides insights into the contribution of clinical metadata to\nthe model’s predictions, supporting both performance gains and\ninterpretability."
 
 RUN useradd -m -u 1000 user
 

README.md

@@ -8,9 +8,12 @@ startup_duration_timeout: 1h
 license: mit
 ---
 
-# Multimodal Skin Lesion Explainability
+# 🔬 RG-DermNet: Multimodal Skin Lesion Explainability
 
 A Gradio-based web application for multimodal skin lesion analysis with **GradCAM++ visualization**. This tool enables clinicians and researchers to understand how deep learning models make predictions on dermoscopic images by combining image data with clinical metadata.
+## 📄 Abstract
+
+Skin cancer accounts for nearly one-third of all diagnosed tumors worldwide, making early and accurate recognition critical for improving patient outcomes. In this work, we propose RG-DermNet, a multimodal deep learning framework that integrates skin lesion images with structured clinical metadata through a residual gated-attention (RG-ATT) fusion mechanism. The architecture combines CNN- and Transformer-based visual backbones with a lightweight one-hot encoding pipeline for metadata, enabling effective cross-modal interaction. The proposed model is evaluated using a patient-wise cross-validation protocol across four dermatological datasets with heterogeneous metadata. On PAD-UFES-20, using Caformer-B36 as the visual backbone, RG-DermNet achieves an accuracy of 0.75 ± 0.05, balanced accuracy of 0.78 ± 0.03, F1-score of 0.77 ± 0.04, and AUC of 0.95 ± 0.01, outperforming existing multimodal baselines under the same evaluation setting. In addition, a SHAP-based analysis provides insights into the contribution of clinical metadata to the model’s predictions, supporting both performance gains and interpretability.
 
 ## 🎯 Features
 
@@ -171,7 +174,7 @@ GradCAMPlusPlus_SkinLesion/
 - **GPU Recommended**: Faster inference with CUDA. Falls back to CPU if unavailable.
 
 ## 🔗 References & Acknowledgments
-
+- **Paper repository**: RG-DermNet (https://github.com/wyctorfogos/rg-dermnet)
 - **Dataset**: PAD-UFES-20 (Universidade Federal do Espírito Santo)
 - **GradCAM++**: [Paper](https://arxiv.org/abs/1710.11063) by Chattopadhyay et al.
 - **Framework**: [Gradio](https://gradio.app) for the web interface
@@ -189,6 +192,16 @@ Contributions are welcome! Please:
 4. Push to the branch (`git push origin feature/improvement`)
 5. Open a Pull Request
 
+## 📎 Citation
+
+@inproceedings{rocha2026rgdermnet,
+  title     = {RG-DermNet: A Multimodal Attention-Based Model with Residual Block Usage for Skin Lesion Classification},
+  author    = {Rocha, Wyctor F. and Bouzon, Pedro H. G. and Ramos, Lucas A. and Pacheco, Andre G. C. and Souza Jr., Luis A.},
+  booktitle = {International Joint Conference on Neural Networks (IJCNN)},
+  year      = {2026},
+  note      = {Accepted}
+}
+
 ## 📧 Support
 
 For issues, questions, or suggestions:

src/main.py

@@ -1,3 +1,5 @@
+import os
+
 import gradio as gr
 from models.inference import (
     run_inference,
@@ -16,26 +18,267 @@ GROUP_CHOICES = [
 DEFAULT_GROUPS = ["demographics", "symptoms", "lesion_geometry"]
 REGION_CHOICES = ["HEAD", "NECK", "BACK", "ARM", "LEG", "TORSO"]
 GENDER_CHOICES = ["MALE", "FEMALE"]
+
 MODEL_CHOICES = get_available_model_choices()
 DEFAULT_MODEL_KEY = get_default_model_key()
 
+PAPER_URL = os.environ.get("PAPER_URL", "").strip()
+PAPER_TITLE = os.environ.get(
+    "PAPER_TITLE",
+    "RG-DermNet: Multimodal Skin Lesion Explainability"
+).strip()
+PAPER_DESCRIPTION = os.environ.get(
+    "PAPER_DESCRIPTION",
+    "This Space accompanies the proposed multimodal framework and allows real-time "
+    "inspection of how clinical metadata affects prediction behavior and GradCAM++ attention maps."
+).strip()
+
+PAPER_FILE = os.environ.get("PAPER_FILE", "paper.pdf").strip()
+PAPER_EXISTS = os.path.exists(PAPER_FILE)
+
 custom_css = """
-.gradio-container { background-color: #0b0f19 !important; }
-.input-panel { border: 1px solid #2d3748; padding: 15px; border-radius: 10px; background: #1a202c; }
-.predict-btn { background: #3182ce !important; color: white !important; font-weight: bold !important; }
-.predict-btn:hover { background: #4299e1 !important; }
-.status-msg { font-style: italic; color: #a0aec0; }
+.gradio-container {
+    background: linear-gradient(180deg, #08111f 0%, #0b0f19 100%) !important;
+    color: #e8eef8 !important;
+}
+
+.main-shell {
+    max-width: 1280px;
+    margin: 0 auto;
+}
+
+.hero {
+    padding: 28px 24px 18px 24px;
+    border: 1px solid #24364d;
+    border-radius: 18px;
+    background: linear-gradient(135deg, rgba(17, 28, 46, 0.95), rgba(8, 13, 24, 0.95));
+    margin-bottom: 18px;
+}
+
+.hero h1 {
+    margin: 0 0 10px 0;
+    font-size: 2.1rem;
+    line-height: 1.2;
+}
+
+.hero p {
+    margin: 0;
+    color: #d2dceb;
+    line-height: 1.65;
+    font-size: 1rem;
+}
+
+.badge-row {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 8px;
+    margin-top: 14px;
+}
+
+.badge {
+    padding: 6px 10px;
+    border-radius: 999px;
+    background: #12253f;
+    border: 1px solid #2f527c;
+    color: #beddff;
+    font-size: 0.9rem;
+}
+
+.section-card {
+    border: 1px solid #24364d;
+    border-radius: 18px;
+    background: rgba(11, 18, 31, 0.92);
+    padding: 18px 20px;
+    margin-bottom: 18px;
+}
+
+.section-card h2,
+.section-card h3 {
+    margin-top: 0;
+}
+
+.paper-card {
+    padding: 16px 18px;
+    border: 1px solid #2d4f7c;
+    border-radius: 14px;
+    background: linear-gradient(135deg, rgba(18, 34, 58, 0.95), rgba(10, 16, 28, 0.95));
+}
+
+.paper-card h3 {
+    margin: 0 0 8px 0;
+    font-size: 1.05rem;
+}
+
+.paper-card p {
+    margin: 0 0 8px 0;
+    color: #d3deee;
+    line-height: 1.55;
+}
+
+.paper-card a {
+    color: #8bc4ff;
+    text-decoration: none;
+    font-weight: 600;
+}
+
+.paper-card a:hover {
+    text-decoration: underline;
+}
+
+.pipeline-box {
+    padding: 14px;
+    border-radius: 14px;
+    border: 1px solid #2b405e;
+    background: #0d1727;
+    text-align: center;
+    min-height: 120px;
+    display: flex;
+    flex-direction: column;
+    justify-content: center;
+}
+
+.pipeline-box h3 {
+    margin-bottom: 8px;
+}
+
+.pipeline-box p {
+    margin: 0;
+    color: #ced9ea;
+    line-height: 1.5;
+}
+
+.pipeline-arrow {
+    text-align: center;
+    font-size: 1.6rem;
+    color: #8bc4ff;
+    padding-top: 38px;
+    font-weight: 700;
+}
+
+.demo-panel {
+    border: 1px solid #2d3748;
+    padding: 16px;
+    border-radius: 16px;
+    background: #111a29;
+}
+
+.predict-btn {
+    background: #3182ce !important;
+    color: white !important;
+    font-weight: bold !important;
+    border: none !important;
+}
+
+.predict-btn:hover {
+    background: #4299e1 !important;
+}
+
+.soft-text {
+    color: #b9c7da;
+    line-height: 1.6;
+}
+
+.footer-note {
+    font-size: 0.95rem;
+    color: #b7c6db;
+    line-height: 1.6;
+}
+
+.paper-frame-wrap {
+    border: 1px solid #2d4f7c;
+    border-radius: 14px;
+    overflow: hidden;
+    background: #0b1321;
+}
+
+.paper-frame {
+    width: 100%;
+    height: 900px;
+    border: none;
+    background: white;
+}
+
+.muted-divider {
+    opacity: 0.25;
+    margin: 10px 0 14px 0;
+}
 """
 
+
+def build_paper_card():
+    link_html = ""
+    if PAPER_URL:
+        link_html += (
+            f'<p><a href="{PAPER_URL}" target="_blank" rel="noopener noreferrer">'
+            f'Open paper link</a></p>'
+        )
+
+    if PAPER_EXISTS:
+        link_html += '<p><a href="/file=paper.pdf" target="_blank" rel="noopener noreferrer">Open embedded PDF in new tab</a></p>'
+
+    if not link_html:
+        link_html = "<p>No external paper link configured yet.</p>"
+
+    return f"""
+    <div class="paper-card">
+        <h3>📄 Associated Paper</h3>
+        <p><strong>{PAPER_TITLE}</strong></p>
+        <p>{PAPER_DESCRIPTION}</p>
+        {link_html}
+    </div>
+    """
+
+
+def build_hero():
+    return """
+    <div class="hero">
+        <h1>🔬 Multimodal Skin Lesion Explainability</h1>
+        <p>
+            This interactive scientific demo presents a multimodal skin lesion analysis system that combines
+            clinical images and patient metadata to generate predictions and GradCAM++ explanations.
+            The interface allows real-time inspection of how metadata influences model attention and
+            diagnostic behavior.
+        </p>
+        <div class="badge-row">
+            <span class="badge">Clinical Image + Metadata</span>
+            <span class="badge">Multimodal Attention</span>
+            <span class="badge">GradCAM++ Explainability</span>
+            <span class="badge">Interactive Paper Demo</span>
+        </div>
+    </div>
+    """
+
+
+def build_paper_embed():
+    if not PAPER_EXISTS:
+        return """
+        <div class="paper-card">
+            <h3>Paper preview unavailable</h3>
+            <p>
+                The file <strong>paper.pdf</strong> was not found in the repository root.
+                Add it to enable in-Space preview.
+            </p>
+        </div>
+        """
+
+    return """
+    <div class="paper-frame-wrap">
+        <iframe src="/file=paper.pdf" class="paper-frame"></iframe>
+    </div>
+    """
+
+
 def format_groups(enabled_groups):
     if not enabled_groups:
         return "No metadata group selected."
     label_map = dict(GROUP_CHOICES)
     return " | ".join([label_map.get(g, g) for g in enabled_groups])
 
+
 def safe_bool(value):
     return bool(value)
 
+
 def build_values_dict(age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation):
     return {
         "age": age,
@@ -51,12 +294,14 @@ def build_values_dict(age, gender, region, diameter1, diameter2, itch, grew, hur
         "elevation": safe_bool(elevation),
     }
 
+
 def build_metadata_preview(enabled_groups, age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation):
     values = build_values_dict(age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation)
     metadata_csv = build_metadata_csv(values, enabled_groups)
     groups_text = format_groups(enabled_groups)
     return metadata_csv, groups_text
 
+
 def validate_inputs(image, enabled_groups, age, diameter1, diameter2):
     if image is None:
         raise gr.Error("Please upload a dermoscopic image first.")
@@ -69,6 +314,7 @@ def validate_inputs(image, enabled_groups, age, diameter1, diameter2):
     if diameter2 is None or diameter2 < 0:
         raise gr.Error("Diameter 2 must be a valid non-negative number.")
 
+
 def gradio_predict(image, selected_model_key, enabled_groups, age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation):
     validate_inputs(image, enabled_groups, age, diameter1, diameter2)
     values = build_values_dict(age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation)
@@ -89,6 +335,7 @@ def gradio_predict(image, selected_model_key, enabled_groups, age, gender, regio
     )
     return image, heatmap_img, pretty_prediction, metadata_csv, groups_text
 
+
 def clear_all():
     default_model = DEFAULT_MODEL_KEY
     if default_model is None and MODEL_CHOICES:
@@ -101,75 +348,168 @@ def clear_all():
         "", format_groups(DEFAULT_GROUPS)
     )
 
-with gr.Blocks(title="Skin Lesion Explainability", theme=gr.themes.Default(primary_hue="blue"), css=custom_css) as demo:
-    gr.Markdown(
-        """
-        # 🔬 Multimodal Skin Lesion Explainability
-        Observe how metadata affects the model's focus using **GradCAM++**.
-        """
-    )
 
-    with gr.Row():
-        with gr.Column(scale=1, variant="panel"):
-            gr.Markdown("### 📥 Input Data")
+with gr.Blocks(
+    title="Skin Lesion Explainability",
+    theme=gr.themes.Default(primary_hue="blue"),
+    css=custom_css,
+) as demo:
+    with gr.Column(elem_classes="main-shell"):
+        gr.HTML(build_hero())
 
-            image_input = gr.Image(type="pil", label="Dermoscopic Image", height=320)
+        with gr.Group(elem_classes="section-card"):
+            gr.Markdown("## 📄 Paper and Scientific Context")
+            with gr.Row():
+                with gr.Column(scale=1):
+                    gr.HTML(build_paper_card())
+                with gr.Column(scale=1):
+                    gr.Markdown(
+                        """
+### Scientific Context
 
-            group_selector = gr.CheckboxGroup(
-                choices=GROUP_CHOICES,
-                value=DEFAULT_GROUPS,
-                label="Enable Metadata Groups"
-            )
+This Space accompanies a multimodal deep learning framework for skin lesion analysis.
+It was designed as an interactive companion to the paper, enabling direct exploration of:
 
-            model_selector = gr.Dropdown(
-                choices=MODEL_CHOICES,
-                value=DEFAULT_MODEL_KEY if DEFAULT_MODEL_KEY is not None else None,
-                label="Attention Mechanism Model",
-                info="Choose which pretrained attention mechanism/model to run.",
-            )
+- metadata-aware prediction behavior
+- qualitative GradCAM++ attention patterns
+- differences across pretrained multimodal attention mechanisms
+
+The goal is to make the proposed method easier to inspect, understand, and communicate.
+""",
+                        elem_classes="soft-text",
+                    )
+
+            with gr.Accordion("📘 Full Paper Preview", open=False):
+                gr.HTML(build_paper_embed())
 
-            with gr.Accordion("👤 Demographics", open=True):
-                age = gr.Number(label="Age", value=55, precision=0)
-                with gr.Row():
-                    gender = gr.Dropdown(GENDER_CHOICES, value="FEMALE", label="Gender")
-                    region = gr.Dropdown(REGION_CHOICES, value="NECK", label="Region")
-
-            with gr.Accordion("📏 Lesion Geometry", open=False):
-                with gr.Row():
-                    diameter1 = gr.Number(label="Diameter 1", value=6)
-                    diameter2 = gr.Number(label="Diameter 2", value=5)
-
-            with gr.Accordion("🚩 Symptoms", open=False):
-                with gr.Row():
-                    itch = gr.Checkbox(label="Itch")
-                    grew = gr.Checkbox(label="Grew")
-                    hurt = gr.Checkbox(label="Hurt")
-                with gr.Row():
-                    changed = gr.Checkbox(label="Changed")
-                    bleed = gr.Checkbox(label="Bleed")
-                    elevation = gr.Checkbox(label="Elevation")
+        with gr.Group(elem_classes="section-card"):
+            gr.Markdown("## 🧠 Method Pipeline")
 
             with gr.Row():
-                clear_btn = gr.Button("Clear", variant="secondary")
-                run_btn = gr.Button("Generate GradCAM++", variant="primary", elem_classes="predict-btn")
+                with gr.Column(scale=1):
+                    gr.HTML(
+                        """
+                        <div class="pipeline-box">
+                            <h3>Input</h3>
+                            <p>Clinical image and structured patient metadata are jointly provided to the system.</p>
+                        </div>
+                        """
+                    )
+                with gr.Column(scale=0):
+                    gr.HTML('<div class="pipeline-arrow">→</div>')
+                with gr.Column(scale=1):
+                    gr.HTML(
+                        """
+                        <div class="pipeline-box">
+                            <h3>Multimodal Model</h3>
+                            <p>Attention-based fusion integrates visual evidence with metadata-aware reasoning.</p>
+                        </div>
+                        """
+                    )
+                with gr.Column(scale=0):
+                    gr.HTML('<div class="pipeline-arrow">→</div>')
+                with gr.Column(scale=1):
+                    gr.HTML(
+                        """
+                        <div class="pipeline-box">
+                            <h3>Output</h3>
+                            <p>The model returns a diagnosis, confidence information, and a GradCAM++ attention map.</p>
+                        </div>
+                        """
+                    )
 
-        with gr.Column(scale=2):
-            gr.Markdown("### 📊 Analysis Dashboard")
+        with gr.Group(elem_classes="section-card"):
+            gr.Markdown("## 🧪 Interactive Demonstration")
+            gr.Markdown(
+                "Modify the metadata, choose a model variant, and inspect how the attention map changes.",
+                elem_classes="soft-text",
+            )
 
             with gr.Row():
-                original_img_out = gr.Image(label="Original Lesion", interactive=False)
-                heatmap_out = gr.Image(label="Attention Map (GradCAM++)", interactive=False)
+                with gr.Column(scale=1, elem_classes="demo-panel"):
+                    gr.Markdown("### 📥 Input Data")
+
+                    image_input = gr.Image(type="pil", label="Dermoscopic Image", height=320)
+
+                    group_selector = gr.CheckboxGroup(
+                        choices=GROUP_CHOICES,
+                        value=DEFAULT_GROUPS,
+                        label="Enable Metadata Groups"
+                    )
+
+                    model_selector = gr.Dropdown(
+                        choices=MODEL_CHOICES,
+                        value=DEFAULT_MODEL_KEY if DEFAULT_MODEL_KEY is not None else None,
+                        label="Attention Mechanism Model",
+                        info="Choose which pretrained attention mechanism/model to run.",
+                    )
+
+                    with gr.Accordion("👤 Demographics", open=True):
+                        age = gr.Number(label="Age", value=55, precision=0)
+                        with gr.Row():
+                            gender = gr.Dropdown(GENDER_CHOICES, value="FEMALE", label="Gender")
+                            region = gr.Dropdown(REGION_CHOICES, value="NECK", label="Region")
+
+                    with gr.Accordion("📏 Lesion Geometry", open=False):
+                        with gr.Row():
+                            diameter1 = gr.Number(label="Diameter 1", value=6)
+                            diameter2 = gr.Number(label="Diameter 2", value=5)
 
-            with gr.Group():
-                prediction_out = gr.Markdown(
-                    value="### Prediction Result\n\nRun the model to see the output here."
-                )
+                    with gr.Accordion("🚩 Symptoms", open=False):
+                        with gr.Row():
+                            itch = gr.Checkbox(label="Itch")
+                            grew = gr.Checkbox(label="Grew")
+                            hurt = gr.Checkbox(label="Hurt")
+                        with gr.Row():
+                            changed = gr.Checkbox(label="Changed")
+                            bleed = gr.Checkbox(label="Bleed")
+                            elevation = gr.Checkbox(label="Elevation")
 
-            with gr.Accordion("📋 System Metadata Details", open=False):
-                active_groups_text = gr.Textbox(label="Active Groups", interactive=False)
-                metadata_preview = gr.Textbox(label="Final CSV Input", lines=6, interactive=False)
+                    with gr.Row():
+                        clear_btn = gr.Button("Clear", variant="secondary")
+                        run_btn = gr.Button("Generate GradCAM++", variant="primary", elem_classes="predict-btn")
+
+                with gr.Column(scale=2, elem_classes="demo-panel"):
+                    gr.Markdown("### 📊 Analysis Dashboard")
+
+                    with gr.Row():
+                        original_img_out = gr.Image(label="Original Lesion", interactive=False)
+                        heatmap_out = gr.Image(label="Attention Map (GradCAM++)", interactive=False)
+
+                    with gr.Group():
+                        prediction_out = gr.Markdown(
+                            value="### Prediction Result\n\nRun the model to see the output here."
+                        )
+
+                    with gr.Accordion("📋 System Metadata Details", open=False):
+                        active_groups_text = gr.Textbox(label="Active Groups", interactive=False)
+                        metadata_preview = gr.Textbox(label="Final CSV Input", lines=6, interactive=False)
+
+        with gr.Group(elem_classes="section-card"):
+            gr.Markdown("## 📚 Notes for Readers")
+            gr.Markdown(
+                """
+- This demo is intended as a qualitative companion to the paper.
+- Users can inspect how metadata groups influence model behavior and attention maps.
+- The available models correspond to pretrained multimodal attention-based variants.
+- For best scientific use, this Space should be interpreted together with the associated manuscript.
+""",
+                elem_classes="footer-note",
+            )
+
+            gr.Markdown("## 📎 Citation and Reproducibility")
+            gr.Markdown(
+                """
+If you reference this demo in a paper, thesis, or presentation, cite the associated manuscript
+and include the Hugging Face Space as supplementary interactive material.
+""",
+                elem_classes="footer-note",
+            )
 
-    preview_inputs = [group_selector, age, gender, region, diameter1, diameter2, itch, grew, hurt, changed, bleed, elevation]
+    preview_inputs = [
+        group_selector, age, gender, region, diameter1, diameter2,
+        itch, grew, hurt, changed, bleed, elevation
+    ]
 
     for component in preview_inputs:
         component.change(