Spaces:
Sleeping
Sleeping
Upload 2 files
Browse filesAI-based skin cancer lesion classifier using a Vision Transformer (ViT) model fine-tuned on the HAM10000 dataset. Upload a dermatological image and receive a detailed risk report with probabilities for 7 common lesion types. For educational and research use only.
- main.py +160 -0
- requirements.txt +6 -0
main.py
ADDED
|
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import ViTFeatureExtractor, ViTForImageClassification
|
| 3 |
+
from PIL import Image
|
| 4 |
+
import matplotlib.pyplot as plt
|
| 5 |
+
import numpy as np
|
| 6 |
+
import gradio as gr
|
| 7 |
+
import io
|
| 8 |
+
import base64
|
| 9 |
+
|
| 10 |
+
MODEL_NAME = "ahishamm/vit-base-HAM-10000-sharpened-patch-32"
|
| 11 |
+
feature_extractor = ViTFeatureExtractor.from_pretrained(MODEL_NAME)
|
| 12 |
+
model = ViTForImageClassification.from_pretrained(MODEL_NAME)
|
| 13 |
+
model.eval()
|
| 14 |
+
|
| 15 |
+
CLASSES = [
|
| 16 |
+
"Queratosis actínica / Bowen", # 0
|
| 17 |
+
"Carcinoma células basales", # 1
|
| 18 |
+
"Lesión queratósica benigna", # 2
|
| 19 |
+
"Dermatofibroma", # 3
|
| 20 |
+
"Melanoma maligno", # 4
|
| 21 |
+
"Nevus melanocítico", # 5
|
| 22 |
+
"Lesión vascular" # 6
|
| 23 |
+
]
|
| 24 |
+
RISK_LEVELS = {
|
| 25 |
+
0: {'level': 'Moderado', 'color': '#ffaa00', 'weight': 0.6},
|
| 26 |
+
1: {'level': 'Alto', 'color': '#ff4444', 'weight': 0.8},
|
| 27 |
+
2: {'level': 'Bajo', 'color': '#44ff44', 'weight': 0.1},
|
| 28 |
+
3: {'level': 'Bajo', 'color': '#44ff44', 'weight': 0.1},
|
| 29 |
+
4: {'level': 'Crítico', 'color': '#cc0000', 'weight': 1.0},
|
| 30 |
+
5: {'level': 'Bajo', 'color': '#44ff44', 'weight': 0.1},
|
| 31 |
+
6: {'level': 'Bajo', 'color': '#44ff44', 'weight': 0.1}
|
| 32 |
+
}
|
| 33 |
+
|
| 34 |
+
def analizar_lesion_vit_web(img):
|
| 35 |
+
inputs = feature_extractor(img, return_tensors="pt")
|
| 36 |
+
with torch.no_grad():
|
| 37 |
+
outputs = model(**inputs)
|
| 38 |
+
probs = outputs.logits.softmax(dim=-1).cpu().numpy()[0]
|
| 39 |
+
pred_idx = int(np.argmax(probs))
|
| 40 |
+
pred_clase = CLASSES[pred_idx]
|
| 41 |
+
confianza = probs[pred_idx]
|
| 42 |
+
cancer_risk_score = sum(probs[i] * RISK_LEVELS[i]['weight'] for i in range(7))
|
| 43 |
+
melanoma_risk = probs[4]
|
| 44 |
+
bcc_risk = probs[1]
|
| 45 |
+
precancer_risk = probs[0]
|
| 46 |
+
benign_total = sum(probs[i] for i in [2,3,5,6])
|
| 47 |
+
|
| 48 |
+
colors_bars = [RISK_LEVELS[i]['color'] for i in range(7)]
|
| 49 |
+
fig, ax = plt.subplots(figsize=(8,3))
|
| 50 |
+
ax.bar(CLASSES, probs*100, color=colors_bars)
|
| 51 |
+
ax.set_title("Probabilidad por tipo de lesión")
|
| 52 |
+
ax.set_ylabel("Probabilidad (%)")
|
| 53 |
+
ax.set_xticklabels(CLASSES, rotation=45, ha='right')
|
| 54 |
+
ax.grid(axis='y', alpha=0.2)
|
| 55 |
+
plt.tight_layout()
|
| 56 |
+
buf = io.BytesIO()
|
| 57 |
+
plt.savefig(buf, format="png")
|
| 58 |
+
plt.close(fig)
|
| 59 |
+
buf.seek(0)
|
| 60 |
+
img_bytes = buf.getvalue()
|
| 61 |
+
img_b64 = base64.b64encode(img_bytes).decode("utf-8")
|
| 62 |
+
html_chart = f'<img src="data:image/png;base64,{img_b64}" style="max-width:100%"/>'
|
| 63 |
+
|
| 64 |
+
urgency = ""
|
| 65 |
+
recommendation = ""
|
| 66 |
+
timeframe = ""
|
| 67 |
+
if cancer_risk_score > 0.6:
|
| 68 |
+
urgency = "🚨 <b>CRÍTICO</b>"
|
| 69 |
+
recommendation = "Derivación INMEDIATA a oncología dermatológica"
|
| 70 |
+
timeframe = "En 24-48 horas máximo"
|
| 71 |
+
elif cancer_risk_score > 0.4:
|
| 72 |
+
urgency = "⚠️ <b>ALTO RIESGO</b>"
|
| 73 |
+
recommendation = "Consulta urgente con dermatólogo especialista"
|
| 74 |
+
timeframe = "En 1 semana máximo"
|
| 75 |
+
elif cancer_risk_score > 0.2:
|
| 76 |
+
urgency = "📋 <b>RIESGO MODERADO</b>"
|
| 77 |
+
recommendation = "Evaluación dermatológica programada"
|
| 78 |
+
timeframe = "En 2-4 semanas"
|
| 79 |
+
else:
|
| 80 |
+
urgency = "✅ <b>BAJO RIESGO</b>"
|
| 81 |
+
recommendation = "Seguimiento de rutina"
|
| 82 |
+
timeframe = "En 3-6 meses"
|
| 83 |
+
|
| 84 |
+
informe = f"""
|
| 85 |
+
<div style="font-family:sans-serif; max-width:700px; margin:auto">
|
| 86 |
+
<h2>🏥 INFORME DE ANÁLISIS DERMATOLÓGICO CON IA</h2>
|
| 87 |
+
<b>📊 Diagnóstico principal:</b> {pred_clase}<br>
|
| 88 |
+
<b>🎯 Confianza del modelo:</b> {confianza:.1%}<br>
|
| 89 |
+
<b>📈 Score de riesgo oncológico:</b> {cancer_risk_score:.1%}<br>
|
| 90 |
+
<br>
|
| 91 |
+
<b>⚠️ Análisis de riesgo detallado:</b><br>
|
| 92 |
+
 🔴 Melanoma maligno: {melanoma_risk:.1%}<br>
|
| 93 |
+
 🟠 Carcinoma células basales: {bcc_risk:.1%}<br>
|
| 94 |
+
 🟡 Lesión pre-cancerosa: {precancer_risk:.1%}<br>
|
| 95 |
+
 🟢 Lesiones benignas: {benign_total:.1%}<br>
|
| 96 |
+
<br>
|
| 97 |
+
<b>🩺 Evaluación clínica:</b><br>
|
| 98 |
+
 {urgency}<br>
|
| 99 |
+
 💡 Recomendación: {recommendation}<br>
|
| 100 |
+
 ⏰ Plazo: {timeframe}<br>
|
| 101 |
+
<br>
|
| 102 |
+
<b>🔍 Características analizadas:</b><br>
|
| 103 |
+
"""
|
| 104 |
+
if melanoma_risk > 0.3:
|
| 105 |
+
informe += "• ⚠️ Posibles irregularidades sugestivas de melanoma<br>• 🔍 Asimetría o variación de color detectada<br>"
|
| 106 |
+
if bcc_risk > 0.25:
|
| 107 |
+
informe += "• 🔍 Características compatibles con carcinoma basocelular<br>"
|
| 108 |
+
if precancer_risk > 0.25:
|
| 109 |
+
informe += "• 🔍 Posible queratosis actínica (lesión pre-maligna)<br>"
|
| 110 |
+
if benign_total > 0.6:
|
| 111 |
+
informe += "• ✅ Características predominantemente benignas<br>"
|
| 112 |
+
|
| 113 |
+
informe += "<br><b>📊 Diagnósticos diferenciales (ordenados por probabilidad):</b><ul>"
|
| 114 |
+
sorted_indices = np.argsort(probs)[::-1]
|
| 115 |
+
for i, idx in enumerate(sorted_indices):
|
| 116 |
+
marker = "🎯" if i==0 else f"{i+1}."
|
| 117 |
+
prob_bars = int(probs[idx] * 30)
|
| 118 |
+
bar_visual = "█" * prob_bars + "░" * (30 - prob_bars)
|
| 119 |
+
risk_indicator = ("🔴" if idx in [1,4] else "🟡" if idx==0 else "🟢")
|
| 120 |
+
informe += f"<li>{marker} {risk_indicator} {CLASSES[idx]} — {probs[idx]:.1%} <code>{bar_visual}</code></li>"
|
| 121 |
+
informe += "</ul>"
|
| 122 |
+
|
| 123 |
+
informe += """
|
| 124 |
+
<br>
|
| 125 |
+
<b>📋 Información técnica:</b><br>
|
| 126 |
+
• Modelo: ViT-Base (HAM10000, sharpened)<br>
|
| 127 |
+
• Precisión validada: 83.7%<br>
|
| 128 |
+
• Preprocesamiento: Optimizado para dermatoscopia<br>
|
| 129 |
+
• Clases detectadas: 7 tipos de lesiones cutáneas<br>
|
| 130 |
+
<br>
|
| 131 |
+
<b>📝 Notas clínicas:</b><br>
|
| 132 |
+
"""
|
| 133 |
+
if confianza < 0.7:
|
| 134 |
+
informe += "• ⚠️ Confianza moderada - considerar segunda opinión<br>"
|
| 135 |
+
if cancer_risk_score > 0.3:
|
| 136 |
+
informe += "• 🩺 Documentar evolución con fotografías seriadas<br>"
|
| 137 |
+
informe += "• 📏 Regla ABCDE: Evaluar Asimetría, Bordes, Color, Diámetro, Evolución<br>"
|
| 138 |
+
|
| 139 |
+
informe += """
|
| 140 |
+
<br><hr>
|
| 141 |
+
<b>⚕️ ADVERTENCIA MÉDICA</b><br>
|
| 142 |
+
• Este análisis es una HERRAMIENTA DE APOYO diagnóstico.<br>
|
| 143 |
+
• NO sustituye la evaluación clínica de un dermatólogo.<br>
|
| 144 |
+
• Ante cualquier duda, consulte con un profesional médico.<br>
|
| 145 |
+
• La decisión terapéutica final corresponde al médico tratante.<br>
|
| 146 |
+
</div>
|
| 147 |
+
"""
|
| 148 |
+
return informe, html_chart
|
| 149 |
+
|
| 150 |
+
demo = gr.Interface(
|
| 151 |
+
fn=analizar_lesion_vit_web,
|
| 152 |
+
inputs=gr.Image(type="pil", label="Sube una imagen de la lesión"),
|
| 153 |
+
outputs=[gr.HTML(label="Informe detallado"), gr.HTML(label="Gráfico de barras")],
|
| 154 |
+
title="Detector de Cáncer de Piel IA (HAM10000, ViT)",
|
| 155 |
+
description="Sube una imagen dermatológica y obtén un informe detallado generado por IA (ViT-HAM10000, accuracy ~83.7%)",
|
| 156 |
+
flagging_mode="never"
|
| 157 |
+
)
|
| 158 |
+
|
| 159 |
+
if __name__ == "__main__":
|
| 160 |
+
demo.launch()
|
requirements.txt
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch
|
| 2 |
+
transformers
|
| 3 |
+
pillow
|
| 4 |
+
matplotlib
|
| 5 |
+
numpy
|
| 6 |
+
gradio
|