{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidlfpjcwgjzgcce6hkhkbqw763ghckyos4aubikdjhpgg6jlodzvi",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mk4vifbq6pp2"
  },
  "path": "/t/medgemma-1-5-4b-useful/175445#post_4",
  "publishedAt": "2026-04-23T00:25:12.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "ISIC",
    "GitHub",
    "ISIC Forum",
    "Kaggle",
    "ISIC Archive Workshop",
    "MICCAI",
    "MLCommons",
    "Nature",
    "ISIC 2024 Challenge Dataset",
    "Google for Developers"
  ],
  "textContent": "Hmm… It doesn’t seem like there’s a single place where all those people gather, but there are a few promising spots:\n\n* * *\n\nFor **your case** , I think the best answer is:\n\n> The serious melanoma / skin-cancer ML people are **mostly not** in one giant always-on forum. They are spread across **ISIC** , **Kaggle ISIC challenge discussions** , **the ISIC Workshop / MICCAI medical-imaging ecosystem** , and a smaller number of **GitHub / benchmark / working-group communities**. Those are the places where the real signal is. (ISIC)\n\n## My read on your situation\n\nYour project is already in a serious zone. Your repo describes an **EVA-02 Small dermoscopy classifier** for **binary malignant vs benign screening** , trained on a curated multi-source ISIC-style dataset, with a held-out test set of **6,384 images** and reported operating points including about **97.01% sensitivity / 82.56% specificity** and **95.02% sensitivity / 88.46% specificity** depending on threshold. That is not a toy result. It means you are already past the “can I train a decent lesion classifier?” stage and into the harder stage: **does it generalize, calibrate, and hold up in tougher settings?** (GitHub)\n\nThat is also why MedGemma likely felt disappointing. You were comparing a **focused dermoscopy screening model** against a **broad medical multimodal foundation model**. Those are different tools. If your goal is “how risky is this mole?”, your closest comparators are not general medical chat models; they are specialized systems like **ADAE** , dermatology foundation models like **Derm Foundation** , and newer dermatology foundation-model work like **PanDerm**. (GitHub)\n\n## Where people actually gather\n\n### 1. ISIC is the closest thing to a home base\n\nIf you only follow one ecosystem, follow **ISIC**. Its AI Working Group explicitly says it is focused on systems for **melanoma diagnosis and clinical decision support** , and it says its AI challenge ecosystem has inspired **over 1,000 scientific papers**. The public ISIC forum has categories for **Events & Workshops**, **ISIC Challenge announcements** , **ISIC Archive** , and **ISIC Live Challenge** , which makes it the closest thing to a dedicated skin-imaging AI forum. (ISIC)\n\nWhat ISIC is good for:\n\n  * official challenge announcements\n  * archive / dataset questions\n  * benchmark infrastructure\n  * finding who is active in this exact niche. (ISIC Forum)\n\n\n\nWhat it is not:\n\n  * it is not a giant informal chat room\n  * it is more structured and official than Discord-like. (ISIC Forum)\n\n\n\n### 2. Kaggle ISIC discussions are where practitioners expose their methods\n\nFor seeing what strong builders are doing **right now** , Kaggle is one of the best places. The **ISIC 2024 Skin Cancer Detection with 3D-TBP** competition focuses on identifying malignant lesions from crops extracted from **3D total-body photography** , and the official dataset site says the training set is **400,000 skin-lesion crops**. Competition discussions and linked solution writeups are where you tend to see:\n\n  * baseline comparisons\n  * leaderboard ablations\n  * model-family choices\n  * augmentation and preprocessing tricks\n  * metric interpretation\n  * code links. (Kaggle)\n\n\n\nFor your project, this matters because Kaggle gives you the most practical view of the field’s current engineering frontier, especially if you want to know “what are good people actually trying?” rather than just “what is published?” (Kaggle)\n\n### 3. The ISIC Workshop at MICCAI is where the serious research conversation happens\n\nThe **ISIC Workshop @ MICCAI** is one of the best places to monitor if you want to know what advanced skin-image-analysis researchers care about. The workshop explicitly positions itself as a venue for advancing skin-image analysis, and the 2024 workshop proceedings include topics like multimodal diagnosis, incomplete multimodal learning, and clinically relevant image-analysis problems. The MICCAI program book also places ISIC within the broader MICCAI ecosystem, which is one of the main global communities for medical image computing. (ISIC Archive Workshop)\n\nThis matters because your next problems are likely to be the same ones they are discussing:\n\n  * generalization\n  * uncertainty\n  * multimodality\n  * class imbalance\n  * fairness\n  * clinically useful risk stratification. (ISIC Archive Workshop)\n\n\n\n### 4. Broader but still relevant: MICCAI and MLCommons Medical AI\n\nIf you want a broader umbrella beyond dermatology-specific spaces, **MICCAI** itself is worth following; the society describes itself as attracting world-leading scientists, engineers, and clinicians in medical image computing. Also, **MLCommons Medical AI** is useful as a broader open benchmark / best-practices community; its working-group page explicitly says it develops benchmarks and best practices for healthcare AI and lists **GitHub** and **Discord** as ways to connect. (MICCAI)\n\nThat said, for **melanoma ML specifically** , I would treat MLCommons as a useful broad-medical-AI side community, not the main place where skin-cancer specialists gather. The highest signal for your problem still seems to be **ISIC + Kaggle + MICCAI workshop activity**. (ISIC)\n\n### 5. Discords: my honest view\n\nI did **not** find a clearly dominant, high-signal, melanoma-specific Discord that I would put on the same level as the communities above. The strongest concrete Discord-style lead I found was the **MLCommons Medical AI** working group, which explicitly lists a Discord, but that is a broad healthcare-AI community rather than a skin-cancer-ML hub. So if you want the best return on time, I would put **formal challenge and workshop ecosystems first** , and Discord second. (MLCommons)\n\n## Where the bar is for your kind of problem\n\nThis is the part that matters most.\n\n### Bar 1: strong retrospective dermoscopy model\n\nYou are already around this bar. Your current repo is a legitimate retrospective dermoscopy screening model with solid reported operating points. This is good work. Many people never get this far. (GitHub)\n\n### Bar 2: external and prospective robustness\n\nThis is the harder bar, and it is the one that separates strong benchmark work from stronger translational work. The **PROVE-AI** prospective validation of **ADAE** reported **96.8% sensitivity** but only **37.4% specificity** at the prespecified 95%-sensitivity threshold. A separate **prospective multicenter** study tested an AI tool on data from **eight hospitals** with different cameras and rare melanoma types, highlighting exactly the kind of heterogeneity that matters in real practice. (Nature)\n\nThat is the background for my main point: your current 97/82-type operating point is promising, but the next important question is not “can I get 0.5% more AUROC?” It is “how much of this survives **site shift, patient shift, camera shift, subtype shift, and prospective workflow constraints**?” (GitHub)\n\n### Bar 3: triage / risk-surfacing workflows\n\nThe field is also moving beyond single-lesion dermoscopy classification. **ISIC 2024** is framed around malignant-lesion identification from **3D total-body-photography crops** , and a follow-up paper reports top ISIC’24 model performance for melanoma classification as high as **AUC 0.9704** with **SE top-15 = 0.7908**. That is a different style of benchmark: closer to **triage and lesion surfacing** than classic dermoscopy-only binary classification. (ISIC 2024 Challenge Dataset)\n\nSo there are really multiple bars:\n\n  * strong retrospective dermoscopy\n  * external / prospective dermoscopy\n  * teletriage / TBP lesion surfacing. (Nature)\n\n\n\n## My thoughts for your case specifically\n\nI think your next best move is **not** to spend too much time trying to make MedGemma into a lesion-risk model. The more relevant comparison class for you is:\n\n  * **ADAE** as a strong public melanoma baseline\n  * **Derm Foundation** as a dermatology-specific embedding model for downstream classifiers\n  * **PanDerm** as a modern dermatology foundation-model direction. (GitHub)\n\n\n\nWhy I say that:\n\n  * **ADAE** is directly in your problem family: melanoma-focused, dermoscopy-oriented, and prospectively validated. (GitHub)\n  * **Derm Foundation** is explicitly designed to accelerate dermatology-image model building with embeddings, requiring less data and compute than training from scratch. That is a practical upgrade path for a project like yours. (Google for Developers)\n  * **PanDerm** is one of the strongest signs of where the field is going: a multimodal dermatology foundation model pretrained on **over 2 million** real-world images from **11 institutions** across **4 modalities** , evaluated on **28 benchmarks** including skin-cancer screening and risk stratification. (Nature)\n\n\n\nThat trio is much closer to your actual problem than a general medical VLM. (GitHub)\n\n## The practical “where should I go?” answer\n\nIf I were in your position, I would do this:\n\n### Check weekly\n\n  * **ISIC forum** for announcements, archive issues, and challenge activity. (ISIC Forum)\n  * **Kaggle ISIC challenge discussions** for practical methods and writeups. (Kaggle)\n  * **ISIC-Research GitHub org** for official code, metrics, and challenge repos. The org exposes repos like **ADAE** and **2024-challenge-dataset**. (GitHub)\n\n\n\n### Check monthly\n\n  * **ISIC Workshop @ MICCAI** papers and proceedings. (ISIC Archive Workshop)\n  * **Derm Foundation** updates and notebooks. (Google for Developers)\n  * **PanDerm** paper/repo ecosystem. (Nature)\n\n\n\n### Join as a side channel\n\n  * **MLCommons Medical AI** if you want a broader med-AI working group with Discord and benchmarks. (MLCommons)\n\n\n\n## My direct recommendation\n\nFor **your exact project** , I would anchor myself in this order:\n\n  1. **ISIC ecosystem first**\n  2. **Kaggle ISIC challenge discussions second**\n  3. **MICCAI / ISIC workshop third**\n  4. **Derm Foundation / ADAE / PanDerm as benchmark and model-family references**\n  5. **MLCommons Medical AI as a broader side community**. (ISIC)\n\n\n\nThat mix will tell you:\n\n  * what others are doing,\n  * what the bar is,\n  * and which next baselines actually matter for your model. (Nature)\n\n",
  "title": "Medgemma 1.5 4b, useful?"
}